<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: helpcode.ai</title>
    <description>The latest articles on DEV Community by helpcode.ai (@helpcode-ai).</description>
    <link>https://dev.to/helpcode-ai</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4249216%2Fcf611e7d-6b47-4cb3-8c55-15bb8cab20ac.png</url>
      <title>DEV Community: helpcode.ai</title>
      <link>https://dev.to/helpcode-ai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/helpcode-ai"/>
    <language>en</language>
    <item>
      <title>MCP code mode in AnythingMCP: context engineering with one program instead of dozens of tool calls</title>
      <dc:creator>helpcode.ai</dc:creator>
      <pubDate>Tue, 17 Nov 2026 13:28:10 +0000</pubDate>
      <link>https://dev.to/helpcode-ai/mcp-code-mode-in-anythingmcp-context-engineering-with-one-program-instead-of-dozens-of-tool-calls-bcj</link>
      <guid>https://dev.to/helpcode-ai/mcp-code-mode-in-anythingmcp-context-engineering-with-one-program-instead-of-dozens-of-tool-calls-bcj</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;In short&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;When an AI agent calls tools one at a time, every intermediate result passes through the model, even when the answer needs five lines of it.&lt;/li&gt;
&lt;li&gt;Code mode lets the agent run one short JavaScript program that makes those calls, filters and joins the results, and returns only the answer.&lt;/li&gt;
&lt;li&gt;In AnythingMCP the program runs in a separate sandbox with no network, no files and no secrets. It can only call the tools the caller may already use.&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  The context window is the budget
&lt;/h2&gt;

&lt;p&gt;Context engineering is the work of deciding what reaches the model. Anthropic defines it as "the set of strategies for curating and maintaining the optimal set of tokens" and adds that context "must be treated as a finite resource with diminishing marginal returns" (&lt;a href="https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents" rel="noopener noreferrer"&gt;Effective context engineering for AI agents&lt;/a&gt;, September 2025).&lt;/p&gt;

&lt;p&gt;Tools are where that budget goes first. MCP, the Model Context Protocol, is the open standard AI clients such as Claude, ChatGPT, Copilot and Cursor use to call tools. With plain MCP, the agent calls one tool, reads the whole result, decides, and calls the next.&lt;/p&gt;

&lt;p&gt;Take a question a sales lead might ask an assistant connected to the CRM and the ERP:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Which of our open quotes belong to customers who still have an overdue invoice?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;With 40 open quotes, the agent lists the quotes, then looks up the invoices of each customer. That is 41 tool calls. Each one is a turn through the model, and each result stays in the context window.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqv22j4vlomkt3s1o7a75.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqv22j4vlomkt3s1o7a75.png" alt="Two columns for the same question. One tool call per step: 41 turns through the model and 41 results in the context. Code mode: one program, one turn, one result in the context" width="800" height="664"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The cost is not special to one vendor. Anthropic describes the same pattern: "Every intermediate result must pass through the model." In its example, a transcript of a two-hour meeting that is copied from one tool to another adds about 50,000 tokens (&lt;a href="https://www.anthropic.com/engineering/code-execution-with-mcp" rel="noopener noreferrer"&gt;Code execution with MCP&lt;/a&gt;, November 2025).&lt;/p&gt;

&lt;h2&gt;
  
  
  What code mode is
&lt;/h2&gt;

&lt;p&gt;Code mode lets an AI client run a short JavaScript program that calls several MCP tools in one step. The program filters, joins or loops over the results, and only its return value goes back to the model.&lt;/p&gt;

&lt;p&gt;The reason it works is simple. Cloudflare put it this way in the post that carries the name: "LLMs are better at writing code to call MCP, than at calling MCP directly" (&lt;a href="https://blog.cloudflare.com/code-mode/" rel="noopener noreferrer"&gt;Code Mode: the better way to use MCP&lt;/a&gt;, September 2025).&lt;/p&gt;

&lt;p&gt;&lt;a href="https://anythingmcp.com" rel="noopener noreferrer"&gt;AnythingMCP&lt;/a&gt; is an open-source MCP gateway. It puts one governed endpoint in front of the ERP, databases and APIs a company already runs. Inside a code-mode program you have exactly one capability, an object called &lt;code&gt;amcp&lt;/code&gt;:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Call&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;amcp.search(query)&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Finds tools the caller may use.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;amcp.describe(tool)&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Returns a tool's input schema.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;amcp.run(tool, args)&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Calls the tool and returns its result as a value.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The 41 calls from the example become this. The tool names are illustrative. In practice they are the tools of your own connectors.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;quotes&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;amcp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'crm_list_open_quotes'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;overdue&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="n"&gt;quotes&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;invoices&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;amcp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'erp_open_invoices'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;customer&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;customerId&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;late&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;invoices&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;filter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;dueDate&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;slice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;late&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;overdue&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;push&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;quote&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kt"&gt;number&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;customer&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;customerName&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;overdueInvoices&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;late&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;overdue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The 40 quote records and the 40 invoice lists never enter the context. The returned list does.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;One tool call per step&lt;/th&gt;
&lt;th&gt;Code mode&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Turns through the model&lt;/td&gt;
&lt;td&gt;One per call (41 in the example)&lt;/td&gt;
&lt;td&gt;One&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;What enters the context&lt;/td&gt;
&lt;td&gt;Every intermediate result&lt;/td&gt;
&lt;td&gt;The program and its return value&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Loops, filters, joins&lt;/td&gt;
&lt;td&gt;Done by the model, by reading&lt;/td&gt;
&lt;td&gt;Done by code, exactly&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best for&lt;/td&gt;
&lt;td&gt;Single lookups, steps that need judgement&lt;/td&gt;
&lt;td&gt;Repetitive, mechanical multi-step work&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuwh13o7trfszy5o2xbkh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuwh13o7trfszy5o2xbkh.png" alt="Two charts. Turns through the model for the worked example: 41 with one tool call per step, 1 with code mode. Tokens spent on tool definitions in Anthropic's example: 150,000 loaded up front, 2,000 read on demand" width="800" height="544"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The second chart is Anthropic's own example, about tool definitions that are read on demand instead of loaded up front.&lt;/p&gt;

&lt;h3&gt;
  
  
  Measured: 20 calls or one program
&lt;/h3&gt;

&lt;p&gt;We ran one task both ways on a self-hosted AnythingMCP 0.55.2 with the public OpenPLZ connector: find the localities of 20 German postal codes and group them by federal state.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;One tool call per postal code&lt;/th&gt;
&lt;th&gt;One saved code tool&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Tool calls the AI client makes&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Characters of tool results that reach the model&lt;/td&gt;
&lt;td&gt;7,283&lt;/td&gt;
&lt;td&gt;895&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That is 88% less text in the context, and one turn through the model in place of twenty. The program is the one from the tutorial below. The saving depends on the task: it grows with the number of calls and with the size of what each call returns.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fvcp1in0hoouuuessx76x.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fvcp1in0hoouuuessx76x.png" alt="Two charts from our own measurement: 20 turns through the model with one tool call per postal code against 1 with a saved code tool, and 7,283 characters of tool results in the context against 895" width="800" height="420"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Three ways to use it
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Saved code tool&lt;/th&gt;
&lt;th&gt;Inline code&lt;/th&gt;
&lt;th&gt;Proposal&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;What it is&lt;/td&gt;
&lt;td&gt;A stored program that appears as an ordinary MCP tool&lt;/td&gt;
&lt;td&gt;The tool &lt;code&gt;code_execute&lt;/code&gt;, which runs a program the agent writes on the spot&lt;/td&gt;
&lt;td&gt;The tool &lt;code&gt;code_propose_tool&lt;/code&gt;, which records a program for an admin to review&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Who writes the program&lt;/td&gt;
&lt;td&gt;An admin, in the dashboard&lt;/td&gt;
&lt;td&gt;The agent&lt;/td&gt;
&lt;td&gt;The agent&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Permissions&lt;/td&gt;
&lt;td&gt;
&lt;em&gt;Invoker&lt;/em&gt;: the caller's own. &lt;em&gt;Definer&lt;/em&gt;: the author's, admin-only&lt;/td&gt;
&lt;td&gt;Always the caller's own&lt;/td&gt;
&lt;td&gt;Nothing runs&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;May change data&lt;/td&gt;
&lt;td&gt;Only with &lt;em&gt;Allow write tools&lt;/em&gt;, on Enterprise&lt;/td&gt;
&lt;td&gt;Never&lt;/td&gt;
&lt;td&gt;Never&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Plans&lt;/td&gt;
&lt;td&gt;Community 3 tools, Starter 5, Team 20, Enterprise unlimited&lt;/td&gt;
&lt;td&gt;Team and Enterprise&lt;/td&gt;
&lt;td&gt;Team and Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A proposal does not run code and does not create a tool. An admin reads the program and clicks &lt;strong&gt;Create tool&lt;/strong&gt;. The new tool starts read-only and runs with each caller's own permissions, so an agent can never raise its own authority.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tutorial: your first code tool
&lt;/h2&gt;

&lt;p&gt;Code mode is off by default everywhere. It needs two things: a running code-runner, and code mode turned on for one MCP server. An MCP server in AnythingMCP is one endpoint with its own set of connectors.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: start the code-runner
&lt;/h3&gt;

&lt;p&gt;On AnythingMCP Cloud the runner is managed for you. Go to step 2.&lt;/p&gt;

&lt;p&gt;On a self-hosted instance (v0.52.1 or later), add two lines to &lt;code&gt;.env&lt;/code&gt; and start the service:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;COMPOSE_PROFILES=code-runner
CODE_RUNNER_URL=ws://code-runner:8777
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;docker compose up -d
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you already use profiles, list them together, for example &lt;code&gt;COMPOSE_PROFILES=files,code-runner&lt;/code&gt;. The runner holds no secrets and needs no inbound internet. Without it, every code run is refused. Code never falls back to running in the backend.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: turn code mode on for one MCP server
&lt;/h3&gt;

&lt;p&gt;In the dashboard, open &lt;strong&gt;MCP Servers&lt;/strong&gt; and then the server. In the &lt;strong&gt;Server settings&lt;/strong&gt; card, set &lt;strong&gt;Code mode&lt;/strong&gt; and click &lt;strong&gt;Save&lt;/strong&gt;.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Setting&lt;/th&gt;
&lt;th&gt;API value&lt;/th&gt;
&lt;th&gt;What it serves&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Off (no code execution)&lt;/td&gt;
&lt;td&gt;&lt;code&gt;off&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Nothing. This is the default.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Saved code tools only&lt;/td&gt;
&lt;td&gt;&lt;code&gt;tools&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Saved code tools&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Saved code tools + inline code&lt;/td&gt;
&lt;td&gt;&lt;code&gt;inline&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Also &lt;code&gt;code_execute&lt;/code&gt; and &lt;code&gt;code_propose_tool&lt;/code&gt;, for roles with the inline-code capability&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The same from the terminal, with the &lt;code&gt;amcp&lt;/code&gt; command line tool and an admin key:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm i -g @anythingmcp/cli
amcp login --browser --scope admin --url https://cloud.anythingmcp.com
amcp api PUT /api/mcp-servers/&lt;span class="nv"&gt;$SRV&lt;/span&gt; --yes --data &lt;span class="s1"&gt;'{"codeMode":"inline"}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Reconnect your AI client afterwards so it sees the new tools.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: create a saved code tool
&lt;/h3&gt;

&lt;p&gt;Open a connector, go to &lt;strong&gt;MCP Tools&lt;/strong&gt; and click &lt;strong&gt;Add Tool&lt;/strong&gt;. Set &lt;strong&gt;Tool Type&lt;/strong&gt; to &lt;strong&gt;Code Tool&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh1jn1k591ph1f0irwirp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh1jn1k591ph1f0irwirp.png" alt="The Add Tool form with Tool Type set to Code Tool: the Program field with JavaScript that loops over postal codes and calls amcp.run, the Uses (connectors) JSON, Run as Invoker, Allow write tools unticked, and one input parameter named postalCodes" width="800" height="965"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Field&lt;/th&gt;
&lt;th&gt;What to enter&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Tool Name&lt;/td&gt;
&lt;td&gt;&lt;code&gt;plz_localities_by_state&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Description&lt;/td&gt;
&lt;td&gt;What the tool returns. The AI reads it to decide when to call the tool.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Program&lt;/td&gt;
&lt;td&gt;The body of an async JavaScript function. Inputs are in &lt;code&gt;args&lt;/code&gt;. Whatever you &lt;code&gt;return&lt;/code&gt; is what the AI receives.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Uses (connectors)&lt;/td&gt;
&lt;td&gt;The connectors, and optionally the exact tools, the program may call. Only connectors of the same workspace.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Run as&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;Invoker (the caller's own permissions)&lt;/strong&gt;: every inner call is checked against the role of the person using the tool.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Allow write tools&lt;/td&gt;
&lt;td&gt;Leave it off. The program may then call only read-only tools.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Input Parameters&lt;/td&gt;
&lt;td&gt;One parameter, &lt;code&gt;postalCodes&lt;/code&gt;, of type string.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;This is the program from the screenshot, and the one we measured above. It uses the OpenPLZ connector from the catalog, which needs no API key, so you can repeat it on your own instance.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;codes&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;postalCodes&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;","&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;map&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;trim&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;filter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;Boolean&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;byState&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;postalCode&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="n"&gt;codes&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;rows&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;amcp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"openplz_lookup_postalcode"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;postalCode&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;row&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="n"&gt;rows&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;state&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;row&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;federalState&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;byState&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;||=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;push&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;postalCode&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="s2"&gt;" "&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;row&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;codes&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;codes&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;byState&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Uses (connectors)&lt;/strong&gt; is a JSON list. A connector's id is the last part of its address, &lt;code&gt;/connectors/&amp;lt;id&amp;gt;&lt;/code&gt;. Add a &lt;code&gt;tools&lt;/code&gt; list to an entry to allow only some of the connector's tools.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"connectorId"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&amp;lt;OpenPLZ connector id&amp;gt;"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A call to a tool outside this list fails, even if the user could call that tool directly.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: test it, then ask
&lt;/h3&gt;

&lt;p&gt;Click the tool's &lt;strong&gt;Test&lt;/strong&gt; button first. It runs the program through the code-runner under the same rules as a real call.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwgesb5jxip8pw5r33w4e.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwgesb5jxip8pw5r33w4e.png" alt="The tool list of the OpenPLZ connector with the new code tool plz_localities_by_state: the Test panel shows the input with five postal codes and the response, Success in 1142 ms, with the localities grouped by federal state" width="800" height="852"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Then refresh the tools in your AI client and ask a question the tool answers. The assistant calls it once.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: run a program from the terminal
&lt;/h3&gt;

&lt;p&gt;With code mode set to &lt;code&gt;inline&lt;/code&gt;, the server offers &lt;code&gt;code_execute&lt;/code&gt;. Log in with an MCP key for that server, save a program to a file and send it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;amcp login --url https://cloud.anythingmcp.com --server SERVER_ID
jq -Rs &lt;span class="s1"&gt;'{code: .}'&lt;/span&gt; program.js &lt;span class="o"&gt;|&lt;/span&gt; amcp tools call code_execute --args -
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Inline code is always read-only and runs with your own permissions.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: let the agent propose a tool
&lt;/h3&gt;

&lt;p&gt;When an agent keeps writing the same program, it can call &lt;code&gt;code_propose_tool&lt;/code&gt;. The proposal appears for admins under &lt;strong&gt;AI Skills&lt;/strong&gt;, in the &lt;strong&gt;Suggested&lt;/strong&gt; tab.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fy6t9s6ijh0txo0htgtpj.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fy6t9s6ijh0txo0htgtpj.png" alt="The AI Skills page in a demo workspace with one suggested item, revenue_by_month, marked Code tool proposal: proposed by an agent, read-only, runs with the caller's permissions, with the reason, the description, the full program and the buttons Create tool and Dismiss" width="800" height="416"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Read the program as you would a small pull request: which tools it calls, what it returns, whether its loops end. Then click &lt;strong&gt;Create tool&lt;/strong&gt; or &lt;strong&gt;Dismiss&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it is safe to run code a model wrote
&lt;/h2&gt;

&lt;p&gt;Code written by a model is untrusted by construction. A prompt-injected tool result can steer what the model writes. So the program never runs in the AnythingMCP backend.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdc86qyqrq7bariqs1t49.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdc86qyqrq7bariqs1t49.png" alt="Architecture: the AI client sends the program to the backend, the only part with secrets, database and internet access. The backend talks over an internal WebSocket to the code-runner container, where a trusted supervisor starts one unprivileged subprocess per run. Below, four layers: runtime permissions, blocked imports, unprivileged user, the container" width="800" height="616"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Each run gets its own subprocess, started fresh and stopped at the time limit. The backend answers every &lt;code&gt;amcp&lt;/code&gt; call by running the real tool under the caller's role, exactly as it would for a direct call.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2wpkuro9595i7m93yz6v.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2wpkuro9595i7m93yz6v.png" alt="Two lists. A program can find, describe and call tools through amcp, loop and filter in JavaScript, log and return one value. It cannot open a network connection, read or write files, read environment variables or secrets, start a process, import modules, change data from inline code or call another code tool. Limits of a single run: 10 seconds, 500 tool calls, 16 at once, 1 MB returned, 128 MB memory" width="800" height="656"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repository ships a containment test that runs real programs through the runner. It checks that &lt;code&gt;fetch&lt;/code&gt;, file reads, environment reads, subprocesses and imports all fail, and that a &lt;code&gt;while(true)&lt;/code&gt; loop is stopped by the deadline.&lt;/p&gt;

&lt;p&gt;Every run is recorded, and each tool call a program makes is recorded like any other tool call. Code tools are never served on the shared &lt;code&gt;/mcp&lt;/code&gt; endpoint. They run only on a workspace's own MCP servers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits and plans
&lt;/h2&gt;

&lt;p&gt;A run stops after 10 seconds by default, 500 tool calls (16 in flight at once) or a returned value above 1 MB. A program over a limit fails with a clear error and affects nothing else.&lt;/p&gt;

&lt;p&gt;On AnythingMCP Cloud the runner is shared between workspaces, so two quotas apply on top:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Cloud plan&lt;/th&gt;
&lt;th&gt;Runs per month&lt;/th&gt;
&lt;th&gt;Runs at the same time&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Trial without a card&lt;/td&gt;
&lt;td&gt;500&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Starter&lt;/td&gt;
&lt;td&gt;500&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Team, and a trial with a card&lt;/td&gt;
&lt;td&gt;5,000&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Enterprise&lt;/td&gt;
&lt;td&gt;Unlimited&lt;/td&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Every Cloud trial has the Team code-mode features: inline code and up to 20 saved code tools. A self-hosted instance is not metered by runs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where code mode fits
&lt;/h2&gt;

&lt;p&gt;Code mode pays off when the work is mechanical: the same lookup for 100 articles, a sum across two systems, two long lists matched against each other.&lt;/p&gt;

&lt;p&gt;A single lookup, or a step that needs judgement such as "read this contract and decide what to check next", stays a normal tool call. And a program the agent writes on the spot never changes data: inline code is read-only by design.&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions people ask
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is code mode in MCP?&lt;/strong&gt;&lt;br&gt;
A way for an AI agent to call MCP tools from inside a short program instead of one at a time. The program runs next to the tools and returns only the result. It does not replace MCP: it calls the same tools through the same gateway and permissions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How does code mode reduce token usage?&lt;/strong&gt;&lt;br&gt;
Intermediate results stay in the sandbox. Only the program and its return value pass through the model. A single lookup saves nothing. The saving grows with the number of calls and the size of the data in between.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which language do I write?&lt;/strong&gt;&lt;br&gt;
JavaScript, as the body of an async function. There are no imports and no packages. The program has &lt;code&gt;amcp&lt;/code&gt; and the language itself.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which AI clients can use it?&lt;/strong&gt;&lt;br&gt;
Any client that supports remote MCP servers, including Claude, ChatGPT, Copilot and Cursor. A saved code tool is a normal MCP tool, and &lt;code&gt;code_execute&lt;/code&gt; is a normal tool that takes a program as its argument.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can a program call an external API directly?&lt;/strong&gt;&lt;br&gt;
No. It has no network access. To reach another API, add it as a connector first, then call its tools with &lt;code&gt;amcp.run&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Documentation: &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/code-mode.md" rel="noopener noreferrer"&gt;Code mode&lt;/a&gt; and the &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/deploy/code-runner/README.md" rel="noopener noreferrer"&gt;code-runner containment model&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Command line tool: &lt;a href="https://www.npmjs.com/package/@anythingmcp/cli" rel="noopener noreferrer"&gt;@anythingmcp/cli&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source code: &lt;a href="https://github.com/HelpCode-ai/anythingmcp" rel="noopener noreferrer"&gt;github.com/HelpCode-ai/anythingmcp&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Cloud trial and self-hosted setup: &lt;a href="https://anythingmcp.com/code-mode" rel="noopener noreferrer"&gt;anythingmcp.com/code-mode&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Start with the one question your team asks every week that takes the assistant more than five tool calls. Save it as a code tool.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>tutorial</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Chat with your documents on a NAS or file server: AnythingMCP File Search for Claude, ChatGPT and Copilot</title>
      <dc:creator>helpcode.ai</dc:creator>
      <pubDate>Tue, 17 Nov 2026 12:24:47 +0000</pubDate>
      <link>https://dev.to/helpcode-ai/chat-with-your-documents-on-a-nas-or-file-server-anythingmcp-file-search-for-claude-chatgpt-and-5g4p</link>
      <guid>https://dev.to/helpcode-ai/chat-with-your-documents-on-a-nas-or-file-server-anythingmcp-file-search-for-claude-chatgpt-and-5g4p</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;In short&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Most company documents sit on a file server or NAS that an AI assistant cannot open, and many companies do not want to copy them to a cloud drive.&lt;/li&gt;
&lt;li&gt;AnythingMCP File Search builds a full-text index of a folder on an SFTP, FTPS, WebDAV or S3 server and keeps it in sync. The files stay where they are.&lt;/li&gt;
&lt;li&gt;Claude, ChatGPT, Copilot and Cursor get tools to search, read pages and find exact codes, and they answer with the file, the page and the date.&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  The documents your AI cannot reach
&lt;/h2&gt;

&lt;p&gt;Ask an AI assistant what an article costs in the current price list, and it has nothing to look at. The price list is a PDF in a folder on the NAS in the server room. So are the drawings, the offers and ten years of saved e-mails.&lt;/p&gt;

&lt;p&gt;The usual advice is to move the files to a cloud drive that the assistant can connect to. For many companies that is not where the files are, and not where they are going.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;In the EU, 52.74% of enterprises bought cloud services in 2025, and 71.53% of those store files there. That is about 38% of all enterprises, by our multiplication of the two shares. For the rest, the files are somewhere else: a server, a NAS, people's computers (&lt;a href="https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Cloud_computing_-_statistics_on_the_use_by_enterprises" rel="noopener noreferrer"&gt;Eurostat, enterprises with 10 or more persons, data extracted January 2026&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;In the US, 62% of IT leaders at large enterprises name reducing data risk from AI the top business challenge of unstructured data, and corporate data leaking into AI models is their top concern about generative AI (&lt;a href="https://www.komprise.com/wp-content/uploads/komprise_2026_unstructured_data_management.pdf" rel="noopener noreferrer"&gt;Komprise 2026 State of Unstructured Data Management, 300 IT leaders, October 2025&lt;/a&gt;).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2x1gnccnnz63myk10f3w.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2x1gnccnnz63myk10f3w.png" alt="Two charts: in the EU about 38% of enterprises store files in a paid cloud service and 62% do not; in the US 62% of IT leaders name reducing data risk from AI the top business challenge of unstructured data" width="800" height="485"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;So the question is how to let the assistant answer from those documents without moving them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What File Search is
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://anythingmcp.com" rel="noopener noreferrer"&gt;AnythingMCP&lt;/a&gt; is an MCP gateway. MCP, the Model Context Protocol, is the open standard that lets AI clients such as Claude, ChatGPT, Copilot and Cursor call tools. The gateway puts one governed endpoint in front of the systems a company already runs.&lt;/p&gt;

&lt;p&gt;File Search is one connector type of that gateway. It lets an AI client search inside the documents of a folder on an SFTP, FTP/FTPS, WebDAV or S3 server, or in a network folder mounted into a self-hosted install. AnythingMCP indexes the file names and the text of every page, keeps the index in sync, and gives the AI tools to search, read pages, find exact codes and hand the file to the person.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb7njakrd4fvp90pv0yvn.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb7njakrd4fvp90pv0yvn.png" alt="Six parts: your file server, the files-worker service that connects and syncs, the isolated file-extractor that turns files into text, the index of file names and page text, the MCP tools, and the AI client that answers with file, page and date" width="800" height="528"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;How the sync works, as documented:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Syncing runs in its own service, &lt;code&gt;files-worker&lt;/code&gt;, never in the process that answers the AI. A large first sync does not slow down your AI clients, and search keeps working on the last index while a sync runs.&lt;/li&gt;
&lt;li&gt;New and changed files are found by size and modification time (by ETag on WebDAV and S3). Unchanged content is recognised by its SHA-256 checksum and not read again. Deleted files leave the index.&lt;/li&gt;
&lt;li&gt;A file that cannot be read is retried, and after three failures it stays findable by name until it changes.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The tools the AI gets
&lt;/h2&gt;

&lt;p&gt;Each File Search connector gets its own tools, named with a prefix you choose.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_search&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Full-text search: the files with the best page, a snippet and the date, with filters for folder, type and dates&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_find&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Files by name, typo-tolerant, or by pattern&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_list&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;One folder: subfolders with counts, files with size and date&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_read&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The text of a file, by pages (&lt;code&gt;"1-3,7"&lt;/code&gt;) or in windows of characters&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_grep&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Exact search with two lines of context; spreadsheet rows come with their column names&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_view_page&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;One page of a PDF, or an image file, as a picture the AI can see: drawings, stamps, signatures&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_get_file&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The file itself as a temporary link for members of the workspace (24 hours)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;&amp;lt;prefix&amp;gt;_status&lt;/code&gt;, &lt;code&gt;&amp;lt;prefix&amp;gt;_guide&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Sync state and counts; how to search this folder&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;prefix&amp;gt;_share_file&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;A temporary link, internal or public. Switched off at install&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Every tool except the share tools is read-only. The tools also tell the model that file content is data, and that instructions found inside a file are never followed.&lt;/p&gt;

&lt;p&gt;Here is a real session against a demo folder of fictional supplier documents. Claude asks for permission to run the exact search:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3hfoz9096lrvi73apu7k.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3hfoz9096lrvi73apu7k.jpg" alt="Claude asks to use the grep tool of the File Search connector with the pattern AD-24004" width="799" height="435"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And this is the answer. It found two price lists that disagree, names both files with page and row, and says which one it would trust and why:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyoo1a7e6idocmowb6fs9.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyoo1a7e6idocmowb6fs9.jpg" alt="Claude's answer quotes the row from a PDF price list and the row from an Excel price list, names both file paths, and explains which list is more likely the current one" width="799" height="435"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Tutorial: a Synology NAS over SFTP
&lt;/h2&gt;

&lt;p&gt;The wizard has five steps, and a sixth connects your AI client. They are the same for any SFTP server. The screenshots are from a local test instance with fictional demo data.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Before you start&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Enable SFTP on the NAS: in DSM, &lt;strong&gt;Control Panel → File Services → FTP → SFTP&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Create a DSM user with read-only access to the one shared folder.&lt;/li&gt;
&lt;li&gt;For AnythingMCP Cloud, the NAS has to be reachable from the internet, with the SFTP port forwarded on your router. A self-hosted AnythingMCP in the same network reaches it directly.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Only workspace administrators can set up File Search connectors.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: Source
&lt;/h3&gt;

&lt;p&gt;Open &lt;strong&gt;Connectors → New → File Search&lt;/strong&gt; and use the &lt;strong&gt;Synology NAS&lt;/strong&gt; shortcut, which presets SFTP. Enter the server, the port, the read-only user and the folder to index.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F526cxdz3j1ytpawb8oq4.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F526cxdz3j1ytpawb8oq4.png" alt="The Source step of the File Search wizard: protocol tiles for SFTP, FTP/FTPS, WebDAV, network folder and S3, shortcuts for Nextcloud, Synology, QNAP, Fritz!Box and Hetzner, and fields for name, server, port, user and folder" width="800" height="713"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: Test
&lt;/h3&gt;

&lt;p&gt;AnythingMCP connects, signs in and lists the folder before anything is saved. The result is a checklist: address, reachability, encryption, server identity, login, folder.&lt;/p&gt;

&lt;p&gt;The SFTP host key is shown as a fingerprint. Confirm it once. From then on a different key is refused before the password is sent.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fif6wbl0vll701xc2vlm2.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fif6wbl0vll701xc2vlm2.png" alt="The Test step: six green checks and a box asking to confirm the fingerprint of the server's SSH host key" width="800" height="518"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: Scan
&lt;/h3&gt;

&lt;p&gt;A dry run lists the whole folder without indexing it: files per folder, types, years. Folders that look like archives (&lt;code&gt;Alt&lt;/code&gt;, &lt;code&gt;Old&lt;/code&gt;, &lt;code&gt;Archive&lt;/code&gt;, &lt;code&gt;Backup&lt;/code&gt;) are unticked. A bar compares your selection with your plan while you tick folders, file types or an age limit.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwgf96z0wxm3ul6pvrig9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwgf96z0wxm3ul6pvrig9.png" alt="The Scan step: 508 files in 212 folders, a budget bar showing 340 selected files against a limit of 2,000, a folder tree with checkboxes, and file type filters" width="800" height="750"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Describe
&lt;/h3&gt;

&lt;p&gt;The wizard drafts the connector's instructions and a one-line summary from the scan: naming patterns of the folders, recurring subfolders, file types, years. This text tells the AI what the folder holds and how to search it. Edit it before saving.&lt;/p&gt;

&lt;p&gt;If an AI model is configured on the server, the wizard can draft the text with AI instead. It states exactly what is sent: folder names, file names and the first 2,000 characters of up to 20 sample files, with personal data masked.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: Activate
&lt;/h3&gt;

&lt;p&gt;Choose the tool prefix, the sync interval and the MCP servers that get the connector. The first sync starts at once.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmxeectuhgi3susp1rxl.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmxeectuhgi3susp1rxl.png" alt="The Activate step: tool prefix, sync interval, the MCP server to add the connector to, and the button to activate and start the first sync" width="800" height="397"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The connector's &lt;strong&gt;Files&lt;/strong&gt; panel then shows the state, the counts against the plan, recent syncs and the files that could not be read.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqif8lzp49dmov6d7k69i.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqif8lzp49dmov6d7k69i.png" alt="The Files panel of a File Search connector: 463 of 508 files indexed, last and next sync, the text the AI is told about the folder, recent syncs including one paused sync, and two files that need attention" width="800" height="931"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: Connect your AI client and ask
&lt;/h3&gt;

&lt;p&gt;In &lt;strong&gt;MCP Servers&lt;/strong&gt;, open the server you chose and copy its URL. In Claude, open &lt;strong&gt;Customize → Connectors&lt;/strong&gt;, click &lt;strong&gt;+&lt;/strong&gt;, then &lt;strong&gt;Add custom connector&lt;/strong&gt;, enter the URL, click &lt;strong&gt;Connect&lt;/strong&gt; and sign in to AnythingMCP.&lt;/p&gt;

&lt;p&gt;Then ask about the content of your files, for example: "Give me the latest price list in the Purchasing folder as a download link."&lt;/p&gt;

&lt;h3&gt;
  
  
  Self-hosted: turn File Search on
&lt;/h3&gt;

&lt;p&gt;On a self-hosted install, two extra services do the work. Add the secrets and the &lt;code&gt;files&lt;/code&gt; profile to &lt;code&gt;.env&lt;/code&gt;, then start:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"FILES_WORKER_URL=http://files-worker:8891"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&amp;gt;&lt;/span&gt; .env
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"FILES_WORKER_TOKEN=$(openssl rand -hex 24)"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&amp;gt;&lt;/span&gt; .env
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"FILE_EXTRACTOR_TOKEN=$(openssl rand -hex 24)"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&amp;gt;&lt;/span&gt; .env
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"COMPOSE_PROFILES=files"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&amp;gt;&lt;/span&gt; .env
docker compose up -d
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For a Windows share or an NFS volume, mount it on the host read-only, bind it into the &lt;code&gt;app&lt;/code&gt; and &lt;code&gt;files-worker&lt;/code&gt; containers and set &lt;code&gt;FILE_SOURCE_ROOTS&lt;/code&gt;, for example &lt;code&gt;FILE_SOURCE_ROOTS=/mnt/knowledge&lt;/code&gt;. The wizard's &lt;strong&gt;Network folder&lt;/strong&gt; source then reads from it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What leaves your server, and what does not
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx38foigluo00wjtlqbhq.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx38foigluo00wjtlqbhq.png" alt="Table: the files stay on your server; file names and page text are in the index; matches, page text, a page picture or a link go to the AI client per question; the server password is stored encrypted; a shared file exists only if share links are switched on" width="800" height="560"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The details behind the table:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The files stay.&lt;/strong&gt; The connector lists and reads. It has no tool that writes, renames or deletes on the server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The index holds text, not files.&lt;/strong&gt; On AnythingMCP Cloud it is stored per workspace, and the database enforces the separation with Row-Level Security. Self-hosted, the index stays on your own server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The AI provider gets what a tool returns&lt;/strong&gt; for one question: a list of matches, the text of a page, a picture of a page, a link.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Credentials&lt;/strong&gt; are stored encrypted and decrypted only in &lt;code&gt;files-worker&lt;/code&gt;. The text extractor runs on an internal network with no route out, no database and no secrets.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Share links are off at install.&lt;/strong&gt; A file shared through a link is deleted after the link expires or is revoked.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cloud or self-hosted.&lt;/strong&gt; AnythingMCP Cloud connects to servers it can reach over the internet, with SFTP or WebDAV. A self-hosted install in your own network also reads SMB and NFS shares directly.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Cloud drive, your own RAG pipeline, or File Search
&lt;/h2&gt;

&lt;p&gt;RAG, retrieval-augmented generation, usually means cutting documents into chunks, turning each chunk into an embedding (a vector that stands for its meaning) and retrieving the most similar chunks for a question.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Copy the files to a cloud drive&lt;/th&gt;
&lt;th&gt;Build your own RAG pipeline&lt;/th&gt;
&lt;th&gt;File Search&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Where the files are&lt;/td&gt;
&lt;td&gt;A second copy at the cloud provider&lt;/td&gt;
&lt;td&gt;On your server, with a copy of the text in your vector database&lt;/td&gt;
&lt;td&gt;On your server; the index holds the text&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Keeping current&lt;/td&gt;
&lt;td&gt;A copy or sync job you run&lt;/td&gt;
&lt;td&gt;Your ingestion pipeline&lt;/td&gt;
&lt;td&gt;Automatic sync; only changed files are read&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Exact codes and numbers&lt;/td&gt;
&lt;td&gt;Depends on the drive's search&lt;/td&gt;
&lt;td&gt;Weak spot: similar codes look alike to embeddings&lt;/td&gt;
&lt;td&gt;Literal search, spreadsheet rows with their column names&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Questions phrased differently from the text&lt;/td&gt;
&lt;td&gt;Depends on the drive's search&lt;/td&gt;
&lt;td&gt;Retrieval by meaning&lt;/td&gt;
&lt;td&gt;The AI client tries several wordings; matching is literal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Who sees what&lt;/td&gt;
&lt;td&gt;Permissions rebuilt in the drive&lt;/td&gt;
&lt;td&gt;You build it&lt;/td&gt;
&lt;td&gt;MCP roles; folder permissions per role on Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sources in the answer&lt;/td&gt;
&lt;td&gt;Depends on the connector&lt;/td&gt;
&lt;td&gt;If you build it&lt;/td&gt;
&lt;td&gt;File, page and date on every hit&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Effort&lt;/td&gt;
&lt;td&gt;A migration&lt;/td&gt;
&lt;td&gt;A project: chunking, embeddings, evaluation&lt;/td&gt;
&lt;td&gt;A wizard: source, test, scan, describe, activate&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;File Search matches words and codes exactly, and the AI client rewords and repeats a search on its own when the first try finds nothing. On a file server, people look for an article number, a name or the newest version of a list. For that, a literal match with a file date is the better tool.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources, file types and limits
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fghaywidi5tilbeeyn7le.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fghaywidi5tilbeeyn7le.png" alt="Sources: SFTP, FTPS, WebDAV, S3, network folder (self-hosted only) and plain FTP. File types read as text: PDF, Excel and CSV, Word, PowerPoint, Outlook messages, OpenDocument, EPUB, ASCII DXF, Markdown, JSON, XML and source code. Found by name: images, archives, DWG and binary DXF. With OCR on Enterprise: scanned PDFs" width="800" height="573"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;File Search is included in every plan of AnythingMCP Cloud, the free trial included. Self-hosted, it needs Enterprise, an existing Starter or Team licence key, or the 30-day trial.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Trial (Cloud)&lt;/th&gt;
&lt;th&gt;Starter&lt;/th&gt;
&lt;th&gt;Team&lt;/th&gt;
&lt;th&gt;Enterprise&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;File Search connectors&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;by contract (self-hosted: unlimited)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Indexed files, all connectors together&lt;/td&gt;
&lt;td&gt;2,000&lt;/td&gt;
&lt;td&gt;500&lt;/td&gt;
&lt;td&gt;2,000&lt;/td&gt;
&lt;td&gt;by contract (self-hosted: unlimited)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Largest file (Cloud)&lt;/td&gt;
&lt;td&gt;50 MB&lt;/td&gt;
&lt;td&gt;25 MB&lt;/td&gt;
&lt;td&gt;50 MB&lt;/td&gt;
&lt;td&gt;200 MB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Automatic sync&lt;/td&gt;
&lt;td&gt;daily&lt;/td&gt;
&lt;td&gt;daily&lt;/td&gt;
&lt;td&gt;daily&lt;/td&gt;
&lt;td&gt;down to every 15 minutes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OCR, folder permissions, workspace policy&lt;/td&gt;
&lt;td&gt;no&lt;/td&gt;
&lt;td&gt;no&lt;/td&gt;
&lt;td&gt;no&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;When a sync finds more files than the plan includes, it pauses before downloading anything. The AI keeps searching the existing index and is told that newer files may be missing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions people ask
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Can Claude or ChatGPT read the files on my NAS directly?&lt;/strong&gt;&lt;br&gt;
No. An AI client cannot open an SFTP or SMB share on its own. It needs a tool that searches for it. File Search provides those tools through an MCP server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Are my files uploaded to the AI provider?&lt;/strong&gt;&lt;br&gt;
No. The provider receives what a tool returns for a question: matches, the text of a page, a picture of a page or a link. The files stay on your server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does this work with SharePoint, OneDrive or Google Drive?&lt;/strong&gt;&lt;br&gt;
Those have connectors of their own that search through the drive's API. File Search is for the places without such an API: a NAS, a file server, an SFTP or WebDAV server, an S3 bucket.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is File Search a RAG system with a vector database?&lt;/strong&gt;&lt;br&gt;
No. It is a full-text index with page-level results, and the model retrieves through tools. It uses no embeddings.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I have to open the NAS to the internet?&lt;/strong&gt;&lt;br&gt;
Not with a self-hosted AnythingMCP in the same network: it needs no port forwarding and also reads mounted SMB and NFS shares. AnythingMCP Cloud connects over the internet, so there the SFTP or WebDAV port has to be reachable, with a read-only user.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What happens with scanned PDFs?&lt;/strong&gt;&lt;br&gt;
Without a text layer they are found by name. On Enterprise, with the OCR extractor image, they are read with OCR.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Documentation: &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/connectors/file-search.md" rel="noopener noreferrer"&gt;File Search and File Share&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Guide: &lt;a href="https://anythingmcp.com/guides/chat-with-documents-vs-rag-vs-file-search" rel="noopener noreferrer"&gt;Chat with your documents, RAG or File Search&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source code: &lt;a href="https://github.com/HelpCode-ai/anythingmcp" rel="noopener noreferrer"&gt;github.com/HelpCode-ai/anythingmcp&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Cloud trial and self-hosted setup: &lt;a href="https://anythingmcp.com" rel="noopener noreferrer"&gt;anythingmcp.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Start with one folder that people ask about every day, and a read-only user for it.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>tutorial</category>
      <category>selfhosted</category>
    </item>
    <item>
      <title>Agent harness or MCP gateway? What a company needs beyond the model, with AnythingMCP and 7 alternatives compared</title>
      <dc:creator>helpcode.ai</dc:creator>
      <pubDate>Tue, 17 Nov 2026 12:24:11 +0000</pubDate>
      <link>https://dev.to/helpcode-ai/agent-harness-or-mcp-gateway-what-a-company-needs-beyond-the-model-with-anythingmcp-and-7-11jh</link>
      <guid>https://dev.to/helpcode-ai/agent-harness-or-mcp-gateway-what-a-company-needs-beyond-the-model-with-anythingmcp-and-7-11jh</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;In short&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The model is the replaceable part. What decides whether AI is useful in a company is the software around it: the agent harness and the MCP gateway.&lt;/li&gt;
&lt;li&gt;This article defines both, lists eleven things to check, and compares eight options against them.&lt;/li&gt;
&lt;li&gt;We build AnythingMCP, one of the eight. Every statement about another product comes from that vendor's own pages, linked below, as read on 17 November 2026.&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  The model is the part you can swap
&lt;/h2&gt;

&lt;p&gt;About one company in five now uses AI. In the European Union, 20.0% of enterprises with 10 or more employees used AI technologies in 2025, up from 13.5% a year earlier (&lt;a href="https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2" rel="noopener noreferrer"&gt;Eurostat, 11 December 2025&lt;/a&gt;). In the United States, 19.8% of businesses used AI in a business function as of 3 May 2026, and 37% of firms with at least 250 employees did (&lt;a href="https://www.census.gov/library/stories/2026/05/ai-use-businesses.html" rel="noopener noreferrer"&gt;U.S. Census Bureau, Business Trends and Outlook Survey, 26 May 2026&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdyj7uscc2gd7z40nxgwh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdyj7uscc2gd7z40nxgwh.png" alt="Two figures side by side: 20.0% of EU enterprises with 10 or more employees used AI in 2025, and 19.8% of US businesses used AI in May 2026, 37% among firms with 250 or more employees" width="800" height="464"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Most of that use starts the same way. Someone opens Claude, ChatGPT or Copilot and asks a question. The answer is good until the question needs something the model cannot see: an order in the ERP, a price list on the file server, a rule the company agreed years ago.&lt;/p&gt;

&lt;p&gt;A better model does not fix that. The model only reads what it is given. Everything else is software around the model, and that software has two names you will meet in every product page this year: agent harness and MCP gateway.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an agent harness is
&lt;/h2&gt;

&lt;p&gt;An agent harness is the runtime around a model. It decides when to call the model, what to send it, how to read the answer, and which tools to call next.&lt;/p&gt;

&lt;p&gt;Microsoft's documentation for Copilot Studio puts it this way: "The harness is a runtime that exists between the two: it determines when to call the model, what components to send it, interprets what comes back, and calls the right tools" (&lt;a href="https://learn.microsoft.com/en-us/microsoft-copilot-studio/harnesses-overview" rel="noopener noreferrer"&gt;Harnesses in Copilot Studio&lt;/a&gt;). The same page lists what depends on the harness: creating and editing files, using skills, remembering context and orchestrating tool calls.&lt;/p&gt;

&lt;p&gt;So when you use Claude, ChatGPT, Copilot or Cursor, you are using a model inside a harness. You usually do not choose the harness separately. It comes with the AI client.&lt;/p&gt;

&lt;p&gt;Harness engineering is the work of designing that runtime: which context the model gets, which tools, which checks before and after each step.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an MCP gateway is
&lt;/h2&gt;

&lt;p&gt;MCP, the &lt;a href="https://modelcontextprotocol.io/introduction" rel="noopener noreferrer"&gt;Model Context Protocol&lt;/a&gt;, is the open standard that lets a harness call tools. An MCP server offers tools. An MCP client, the harness, calls them.&lt;/p&gt;

&lt;p&gt;An MCP gateway is one MCP endpoint in front of many tools. It handles sign-in, access control and the record of calls for all of them. AI clients connect to the gateway and not to each system.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6b1g2ahevq8v8pmc136f.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6b1g2ahevq8v8pmc136f.png" alt="Four layers: the model, the agent harness that runs the loop and calls tools, the MCP gateway that checks identity and roles and records each call, and the company systems behind it" width="800" height="704"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The two are not alternatives. The harness sits with the model and belongs to the AI client. The gateway sits with your systems and belongs to you.&lt;/p&gt;

&lt;p&gt;That split matters once a company uses more than one AI client, and most do. Memory, skills and file access are usually counted as harness features. If they live only in one harness, they stay in one AI client. If the gateway serves them, every harness that connects gets the same ones.&lt;/p&gt;

&lt;h2&gt;
  
  
  Eleven things to check beyond the model
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fj90zwabkivervmsqav9m.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fj90zwabkivervmsqav9m.png" alt="A checklist in five groups: reach (connectors, file access), control (identity, roles per tool, audit log), knowledge (memory, skills), execution (code in a sandbox) and terms (hosting, licence, price model)" width="800" height="666"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Reach&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Connectors for the systems you run.&lt;/strong&gt; Count the ones you have, not the size of the catalog. Check the old ones too: a SOAP service, an OData service on SAP, a SQL database that is the real source of truth.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;File access.&lt;/strong&gt; Much of what a company knows is in PDFs and spreadsheets on a file server. Ask whether the AI can search them and say which file, page and date an answer comes from.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F76kufskz47iyqnopbpnb.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F76kufskz47iyqnopbpnb.png" alt="A Claude answer from demo documents that quotes a price row and names the file, the page and the validity date of two price lists" width="680" height="667"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Control&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Identity.&lt;/strong&gt; Single sign-on from your directory, and per-user sign-in to the connected systems, so that each person sees what they are allowed to see there.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Roles per tool.&lt;/strong&gt; Start read-only. The role should decide which tools a person's AI can call.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audit log.&lt;/strong&gt; Who called which tool, with which input, and what came back.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Knowledge&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Memory.&lt;/strong&gt; A fact said once in one AI client should be known in the others.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Skills.&lt;/strong&gt; The way your company does a piece of work, written down as a procedure the AI follows.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Execution&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Code in a sandbox.&lt;/strong&gt; One short program that calls several tools is cheaper and more exact than twenty single calls. It must run isolated, with no network and no secrets.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Terms&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Hosting.&lt;/strong&gt; Your own servers, or a region you choose.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Licence.&lt;/strong&gt; Whether you can read the code and keep running it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Price model.&lt;/strong&gt; Flat, per seat, or per tool call. An agent conversation can make dozens of calls.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Eight options compared
&lt;/h2&gt;

&lt;p&gt;We picked the options a buyer would put on a shortlist today: two gateways from the API and AI infrastructure side, three integration platforms that added MCP, one agent builder with its own harness, one tool platform for agents, and our own.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftjxpn5fysuquem157m3n.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftjxpn5fysuquem157m3n.png" alt="A matrix of eight products against seven criteria: self-hosted, your own APIs as tools, roles and audit log, skills, memory, file search and code sandbox, with marks for stated, partly, not available and not stated" width="800" height="724"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;How to read the marks: a dot means the vendor's page did not say. It does not mean the feature is missing. The three tables below name the plan behind each mark.&lt;/p&gt;

&lt;h3&gt;
  
  
  Hosting, licence and price
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Option&lt;/th&gt;
&lt;th&gt;What it is&lt;/th&gt;
&lt;th&gt;Where it runs&lt;/th&gt;
&lt;th&gt;Licence&lt;/th&gt;
&lt;th&gt;Price model&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;AnythingMCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;MCP gateway that also generates the tools from your APIs&lt;/td&gt;
&lt;td&gt;Your servers with Docker, or AnythingMCP Cloud in Frankfurt&lt;/td&gt;
&lt;td&gt;AGPL-3.0; some modules under a commercial licence&lt;/td&gt;
&lt;td&gt;Self-hosted Community free up to 3 active users; Cloud €19 or €49 a month, flat; Enterprise quoted&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Composio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Tool platform for agents, with an MCP gateway&lt;/td&gt;
&lt;td&gt;Composio Cloud; bring your own cloud on Enterprise&lt;/td&gt;
&lt;td&gt;SDK repository MIT; platform not stated&lt;/td&gt;
&lt;td&gt;Free with 10,000 tool calls a month; Standard $18 per seat a month; extra calls $0.005 each; Enterprise custom&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Zapier MCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Hosted MCP server over Zapier's apps&lt;/td&gt;
&lt;td&gt;Zapier's cloud, data in AWS US-East 1; no VPC or on-premises&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Runs on your existing Zapier plan with no separate MCP billing; each successful tool call uses two tasks&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;n8n&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Workflow automation; workflows you enable become MCP tools&lt;/td&gt;
&lt;td&gt;Self-hosted, or n8n Cloud&lt;/td&gt;
&lt;td&gt;Sustainable Use License; Enterprise files under a separate licence&lt;/td&gt;
&lt;td&gt;Self-hosted Community free; Cloud from €20 a month for 2,500 executions; Business €667 a month; Enterprise quoted&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Workato&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Integration platform with Enterprise MCP&lt;/td&gt;
&lt;td&gt;Workato's cloud; a dedicated AWS VPC as Virtual Private Workato&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Quote only; usage tracked by tasks, on-premises agents and concurrency&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Microsoft Copilot Studio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Agent builder with its own harnesses and an MCP client&lt;/td&gt;
&lt;td&gt;Microsoft's service; self-hosting not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Copilot Credits: $200 a month per 25,000 credits, or pay as you go&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Kong AI Gateway&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;API gateway with MCP plugins&lt;/td&gt;
&lt;td&gt;On-premises, or Kong Konnect&lt;/td&gt;
&lt;td&gt;MCP plugin only in the AI Gateway Enterprise offering&lt;/td&gt;
&lt;td&gt;Konnect Plus from $25 a month plus usage; Enterprise and self-hosted quoted&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;TrueFoundry&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;AI gateway with an MCP gateway and registry&lt;/td&gt;
&lt;td&gt;SaaS; VPC, on-premises and air-gapped on Enterprise&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Free for up to 3 users; Pro $25 per user a month; Enterprise custom&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  Reach and control
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Option&lt;/th&gt;
&lt;th&gt;Your own APIs as tools&lt;/th&gt;
&lt;th&gt;Roles and audit log&lt;/th&gt;
&lt;th&gt;Identity&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;AnythingMCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;800+ ready connectors; REST/OpenAPI, SOAP/WSDL, GraphQL, OData and SQL sources become tools without code&lt;/td&gt;
&lt;td&gt;Roles, per-tool permissions and an audit log in every edition&lt;/td&gt;
&lt;td&gt;SSO and SCIM on self-hosted Enterprise; per-user sign-in on Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Composio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;1,700+ apps; custom connectors for internal apps on Enterprise&lt;/td&gt;
&lt;td&gt;Roles and tool scoping; audit logs on Enterprise&lt;/td&gt;
&lt;td&gt;SSO over SAML and OIDC on Standard and Enterprise; SCIM on Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Zapier MCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;9,000+ apps; your own APIs not stated&lt;/td&gt;
&lt;td&gt;App and action restrictions set for the account; a history of tool calls and an audit log&lt;/td&gt;
&lt;td&gt;SAML SSO on Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;n8n&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Each tool is a workflow you build and enable for MCP&lt;/td&gt;
&lt;td&gt;Custom roles and audit logging on Enterprise&lt;/td&gt;
&lt;td&gt;SSO and LDAP on Business and Enterprise&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Workato&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;12,000+ apps; recipe functions and skills become MCP servers&lt;/td&gt;
&lt;td&gt;Role-based access control and audit trails&lt;/td&gt;
&lt;td&gt;Agent actions inherit the signed-in user's identity&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Microsoft Copilot Studio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Power Platform connectors, custom connectors and MCP servers&lt;/td&gt;
&lt;td&gt;Power Platform data policies also govern MCP servers; audit not stated&lt;/td&gt;
&lt;td&gt;OAuth 2.0 per user for MCP servers&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Kong AI Gateway&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Any Kong-managed API becomes MCP tools from its schema&lt;/td&gt;
&lt;td&gt;Access lists per tool, audit logs and metrics&lt;/td&gt;
&lt;td&gt;OAuth 2.0 plugin for MCP&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;TrueFoundry&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Registered MCP servers; internal APIs through custom servers&lt;/td&gt;
&lt;td&gt;Role-based access per MCP server; audit logging&lt;/td&gt;
&lt;td&gt;OAuth 2.0; SSO with Okta, Azure AD and others&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  Knowledge and execution
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Option&lt;/th&gt;
&lt;th&gt;Skills&lt;/th&gt;
&lt;th&gt;Memory&lt;/th&gt;
&lt;th&gt;File search&lt;/th&gt;
&lt;th&gt;Code sandbox&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;AnythingMCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;AI skills in every edition; learning projects on Team and Enterprise&lt;/td&gt;
&lt;td&gt;Company memory shared by every AI client, on Enterprise&lt;/td&gt;
&lt;td&gt;SFTP, FTP, WebDAV, S3 and network shares; on Cloud, or self-hosted Enterprise&lt;/td&gt;
&lt;td&gt;Code Mode in an isolated runner, switched on per MCP server&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Composio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;"Skill sharing" listed in the plans; details not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Sandboxes "to run code and process files"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Zapier MCP&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;n8n&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Workato&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;"Enterprise Skills": recipe functions served as MCP servers&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Microsoft Copilot Studio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Skills on two of its three harnesses&lt;/td&gt;
&lt;td&gt;Memory on the same two harnesses, for agents built in Copilot Studio&lt;/td&gt;
&lt;td&gt;Creates and edits Word, Excel, PowerPoint and PDF files; file server search not stated&lt;/td&gt;
&lt;td&gt;Each task runs in a sandbox on the Microsoft Copilot harness&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Kong AI Gateway&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;TrueFoundry&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Agent Skills Registry, a separate product&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;td&gt;Agents can "generate and execute code"; sandboxing not described&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Sources, all read on 17 November 2026:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;AnythingMCP:&lt;/strong&gt; &lt;a href="https://github.com/HelpCode-ai/anythingmcp" rel="noopener noreferrer"&gt;README&lt;/a&gt;, &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/cloud-plans.md" rel="noopener noreferrer"&gt;Cloud plans&lt;/a&gt;, &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/license-faq.md" rel="noopener noreferrer"&gt;licence FAQ&lt;/a&gt;, &lt;a href="https://anythingmcp.com/pricing" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Composio:&lt;/strong&gt; &lt;a href="https://composio.dev/pricing" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;, &lt;a href="https://composio.dev/enterprise" rel="noopener noreferrer"&gt;enterprise&lt;/a&gt;, &lt;a href="https://github.com/ComposioHQ/composio/blob/master/LICENSE" rel="noopener noreferrer"&gt;SDK licence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zapier MCP:&lt;/strong&gt; &lt;a href="https://zapier.com/mcp" rel="noopener noreferrer"&gt;product page&lt;/a&gt;, &lt;a href="https://docs.zapier.com/mcp/features/usage" rel="noopener noreferrer"&gt;usage&lt;/a&gt;, &lt;a href="https://docs.zapier.com/mcp/manage/security" rel="noopener noreferrer"&gt;security&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;n8n:&lt;/strong&gt; &lt;a href="https://n8n.io/pricing/" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;, &lt;a href="https://docs.n8n.io/connect/connect-to-n8n-mcp-server" rel="noopener noreferrer"&gt;MCP server docs&lt;/a&gt;, &lt;a href="https://docs.n8n.io/choose-n8n/" rel="noopener noreferrer"&gt;editions&lt;/a&gt;, &lt;a href="https://github.com/n8n-io/n8n/blob/master/LICENSE.md" rel="noopener noreferrer"&gt;licence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Workato:&lt;/strong&gt; &lt;a href="https://www.workato.com/agentic/mcp" rel="noopener noreferrer"&gt;Enterprise MCP&lt;/a&gt;, &lt;a href="https://www.workato.com/pricing" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;, &lt;a href="https://docs.workato.com/en/security/security-compliance" rel="noopener noreferrer"&gt;compliance&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Microsoft Copilot Studio:&lt;/strong&gt; &lt;a href="https://learn.microsoft.com/en-us/microsoft-copilot-studio/harnesses-overview" rel="noopener noreferrer"&gt;harnesses&lt;/a&gt;, &lt;a href="https://learn.microsoft.com/en-us/microsoft-copilot-studio/agent-extend-action-mcp" rel="noopener noreferrer"&gt;MCP&lt;/a&gt;, &lt;a href="https://learn.microsoft.com/en-us/microsoft-copilot-studio/mcp-add-existing-server-to-agent" rel="noopener noreferrer"&gt;connect an MCP server&lt;/a&gt;, &lt;a href="https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kong AI Gateway:&lt;/strong&gt; &lt;a href="https://developer.konghq.com/mcp/" rel="noopener noreferrer"&gt;MCP docs&lt;/a&gt;, &lt;a href="https://developer.konghq.com/plugins/ai-mcp-proxy/" rel="noopener noreferrer"&gt;AI MCP Proxy plugin&lt;/a&gt;, &lt;a href="https://konghq.com/pricing" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;TrueFoundry:&lt;/strong&gt; &lt;a href="https://www.truefoundry.com/mcp-gateway" rel="noopener noreferrer"&gt;MCP Gateway&lt;/a&gt;, &lt;a href="https://www.truefoundry.com/pricing" rel="noopener noreferrer"&gt;pricing&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Where each option fits best
&lt;/h2&gt;

&lt;p&gt;No option is the best fit for every company. Read the tables with your own situation in mind.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;AnythingMCP&lt;/strong&gt; fits a company with more than one AI client and with systems that have no MCP server of their own: an ERP, SOAP or OData services, SQL databases, file servers. It generates the tools from those sources and serves skills, memory and file search from the gateway, so Claude, ChatGPT and Copilot all get the same ones. We did not find that combination stated by any other vendor above.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zapier MCP&lt;/strong&gt; has the largest catalog, 9,000+ apps, and nothing to run. For standard SaaS actions and light use it is the shortest path.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Composio&lt;/strong&gt; covers 1,700+ apps and includes sandboxes that run code. Its paid plans add a firewall and zero data retention.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Workato&lt;/strong&gt; suits a company that already runs it as its integration platform. MCP is then an extension of what is there.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Microsoft Copilot Studio&lt;/strong&gt; brings skills, memory, file editing and a sandbox inside its own harness. For a company that works in Microsoft 365 and stays there, that is a strong reason.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kong AI Gateway&lt;/strong&gt; fits a company that already publishes its APIs through Kong. The same gateway then serves them as MCP tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;n8n&lt;/strong&gt; is the better engine when the AI is one step in a longer workflow with branches, schedules and error handling.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;TrueFoundry&lt;/strong&gt; counts model requests and MCP tool calls on one platform and offers air-gapped deployment on Enterprise.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  A 10-minute self-hosted start
&lt;/h2&gt;

&lt;p&gt;This is the quickstart from the AnythingMCP README. It needs Docker 24 or newer and &lt;code&gt;openssl&lt;/code&gt;. On macOS, start Docker Desktop first.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;mkdir anythingmcp &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;cd&lt;/span&gt; anythingmcp
curl -fsSLo docker-compose.yml \
  https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml
&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n'&lt;/span&gt; &lt;span class="s2"&gt;"$(openssl rand -hex 32)"&lt;/span&gt; &lt;span class="s2"&gt;"$(openssl rand -hex 32)"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; .env
docker compose up -d
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open &lt;code&gt;http://localhost:3000&lt;/code&gt; and register. The first account becomes admin. Your MCP endpoint is &lt;code&gt;http://localhost:4000/mcp&lt;/code&gt;, and the API documentation is at &lt;code&gt;http://localhost:4000/api/docs&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Keep the generated &lt;code&gt;.env&lt;/code&gt; file. &lt;code&gt;ENCRYPTION_KEY&lt;/code&gt; decrypts the credentials you store. If you lose it, every connector needs its credentials again.&lt;/p&gt;

&lt;p&gt;Then three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Add a connector.&lt;/strong&gt; Install one from the catalog, or point AnythingMCP at your own OpenAPI spec, WSDL, GraphQL endpoint, Postman collection or database. Each operation becomes an MCP tool.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decide what the model sees.&lt;/strong&gt; Rename and describe the tools, drop the fields that must not leave your network, and choose which roles may call which tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hand one URL to your AI client.&lt;/strong&gt; An MCP server in AnythingMCP is the endpoint you add to Claude, ChatGPT, Copilot, Gemini or Cursor. It exposes the connectors you assign to it and nothing else.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2x6nn9dn2tyk5owc461r.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2x6nn9dn2tyk5owc461r.png" alt="The tool list of a SAP S/4HANA connector in AnythingMCP: ten MCP tools, each with Test, Edit, Hints, Disable and Delete" width="800" height="511"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The quickstart listens on &lt;code&gt;127.0.0.1&lt;/code&gt; only, because nothing in front of it handles TLS. For an instance that colleagues or a cloud AI client can reach, clone the repository and run &lt;code&gt;./setup.sh&lt;/code&gt;. It asks for a domain, gets certificates and generates the secrets. The &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/deployment.md" rel="noopener noreferrer"&gt;deployment guide&lt;/a&gt; has the details.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to choose
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Your situation&lt;/th&gt;
&lt;th&gt;Look first at&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Several AI clients, and systems with no MCP server: ERP, SOAP, databases, file servers&lt;/td&gt;
&lt;td&gt;AnythingMCP&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Everything must run on your own servers at no licence cost&lt;/td&gt;
&lt;td&gt;AnythingMCP Community or n8n Community&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Standard SaaS apps, light use, nothing to operate&lt;/td&gt;
&lt;td&gt;Zapier MCP&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A large SaaS catalog, priced per seat&lt;/td&gt;
&lt;td&gt;Composio&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;You already run Workato, Kong or n8n&lt;/td&gt;
&lt;td&gt;The MCP feature of the platform you have&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Your company works in Microsoft 365 and will stay there&lt;/td&gt;
&lt;td&gt;Copilot Studio, with a gateway for the systems its connectors do not cover&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;You also need to govern model traffic and cost&lt;/td&gt;
&lt;td&gt;TrueFoundry or Kong AI Gateway&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A gateway and a harness can be combined. Copilot Studio and the common AI clients can call an MCP gateway as one of their tool sources.&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions people ask
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is an agent harness?&lt;/strong&gt;&lt;br&gt;
The runtime around a model. It decides when to call the model, what context to send, and which tools to call with the answer. Claude, ChatGPT, Copilot and Cursor each contain one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is an MCP gateway?&lt;/strong&gt;&lt;br&gt;
One MCP endpoint in front of many tools. It handles sign-in, access control and the record of calls, so AI clients connect to the gateway and not to each system.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Agent harness or MCP gateway: do I need both?&lt;/strong&gt;&lt;br&gt;
You already have a harness if you use an AI client. You need a gateway when more than one person or more than one AI client should reach your systems under rules you set.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is harness engineering?&lt;/strong&gt;&lt;br&gt;
The work of designing a harness: which context the model gets, which tools it can call, and which checks run before and after each step.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My AI client already has connectors. Why add a gateway?&lt;/strong&gt;&lt;br&gt;
Built-in connectors cover well-known SaaS apps and live in one AI client. A gateway reaches the systems that have no ready connector, and gives every AI client the same tools, roles and log.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which of these can I self-host for free?&lt;/strong&gt;&lt;br&gt;
AnythingMCP states a free Community edition for up to 3 active users, and n8n a free Community edition. Composio, Kong and TrueFoundry tie self-hosting to an Enterprise arrangement. Zapier says it is not available.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is AnythingMCP open source?&lt;/strong&gt;&lt;br&gt;
Yes, under AGPL-3.0. Some modules, among them the File Search sync engine and learning projects, are under a commercial licence and can be read in the same repository.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Source code: &lt;a href="https://github.com/HelpCode-ai/anythingmcp" rel="noopener noreferrer"&gt;github.com/HelpCode-ai/anythingmcp&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Documentation: &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/deployment.md" rel="noopener noreferrer"&gt;deployment&lt;/a&gt;, &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/connectors/file-search.md" rel="noopener noreferrer"&gt;File Search&lt;/a&gt;, &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/memory.md" rel="noopener noreferrer"&gt;memory&lt;/a&gt;, &lt;a href="https://github.com/HelpCode-ai/anythingmcp/blob/main/docs/code-mode.md" rel="noopener noreferrer"&gt;Code Mode&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Cloud trial, 7 days, hosted in Frankfurt: &lt;a href="https://anythingmcp.com" rel="noopener noreferrer"&gt;anythingmcp.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you compare options yourself, take the eleven points above to each vendor and ask for the page where each one is stated.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>agents</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
