rambo

Posted on Oct 12

Add Zambo to Your AI: The Complete MCP Setup Guide (Every Client, 2026)

#aiagents #mcp #aiverification #tutorial

Add Zambo to Your AI. Switch Models. Keep the Receipts.

Agents do work.

Then somebody asks what happened, and we hand them a chat transcript.

"Done" is not a receipt. A plan is not execution. A confident answer is not evidence.

That's the gap I'm building Zambo around.

One hosted MCP connection. A record of what Zambo recorded. A link you can inspect instead of asking the model to tell the story again.

What a Receipt Is

A Zambo receipt records the tool call, its status, and committed result bytes. You can retrieve the record and check whether the bytes match the published hash.

That makes changes detectable against the commitment. It does not make the answer true. It does not independently authenticate the issuer. And it does not prove an outside action happened unless there's corresponding evidence of that action.

Wrong result? You can have an intact receipt of a wrong result.

Failed call? That's a failed call, not completed work.

That's the point. Keep the claim smaller than the evidence.

The Workflow Update

A job can have more than one step.

The workflow receipt connects the step receipts into one inspectable workflow record, with a Merkle root over the ordered step commitments. The individual receipts stay intact.

One workflow link to review the run. Individual step links when you need the detail.

The workflow page's verify control fetches the step receipts again, rebuilds the tree in your browser, and compares the root. A match checks that the manifest agrees with those step records. It isn't a magic stamp on every claimed outcome.

workflow_close takes a snapshot of the eligible receipts in a Zambo session. Call it again without changing the job data and you get the same snapshot. Add more work and a later snapshot can grow.

So don't confuse a workflow grouping with a promise that the entire future job is frozen.

The Connection

This is the address:

https://zambo.dev/mcp

No Zambo signup or API key to start. No OAuth for the anonymous endpoint.

Your AI platform can still require its own account, plan, permissions, or API key. Free Zambo does not mean free model usage.

Free limits vary by tool. Don't read "free" as unlimited access to everything. The live catalog and tool contracts tell you the allowance.

Connect one Zambo server per client. Not one server per tool.

The steps below follow the official docs checked October 11, 2026. I checked the source wording, reviewed the config shapes, and checked the examples again against the endpoint's live contract.

The review had three passes: official setup docs, corroborating references and conflict checks, then config syntax and live endpoint checks. I did not connect every app in an authenticated account. Documented support is not a guarantee for every account. Unresolved paths below say unverified.

Claude Desktop and Claude.ai

Use the native remote connector. You don't need a local bridge for the free connection.

For an individual account:

Open Customize > Connectors.
Choose + Add > Add custom connector.
Name it zambo.
Paste https://zambo.dev/mcp and continue.
Choose no sign-in if the dialog asks. Don't invent OAuth credentials.
Finish adding it.
In a chat, use + > Connectors to enable Zambo for that conversation.

Free accounts get one custom connector. Pro and Max can also add one. On team or enterprise, an owner or authorized admin adds the connector under Organization Settings > Connectors > Add > Custom > Web. Members then connect to it from their own connectors list.

Claude Desktop's native remote connection comes from Anthropic's cloud, not your laptop. Don't use a private localhost address and expect this flow to reach it.

ChatGPT

ChatGPT has a real native remote MCP path. GPT Actions are a different thing.

OpenAI's current developer guide says:

On ChatGPT web, go to Plugins.
Select + > Add custom MCP server.
Name it zambo.
Under connection, use https://zambo.dev/mcp as the server URL.
Choose no authentication.
Review the warning, then create as a plugin.
Install it. Type @ in the prompt box and select it.

There's a documentation mismatch worth saying out loud: OpenAI's help center still describes Settings > Apps > Advanced Settings > Developer Mode, then Apps > Create. It limits full MCP to Business, Enterprise, and Edu, while the developer guide also lists Plus and Pro.

If your account shows the apps flow, use that flow. If it shows neither custom-server option, don't assume pasting the URL into chat installed anything. Check your plan and workspace permissions. An admin may need to enable it.

Review tool-call confirmations. Don't globally approve every write action just to make setup feel faster.

Custom GPTs / Actions

A GPT Action consumes an OpenAPI schema. /mcp is a JSON-RPC MCP endpoint, not an OpenAPI schema you can import.

If you're building a custom GPT, the real alternative is an HTTP API wrapper: expose specific operations, call Zambo inside that wrapper, describe them with a tested OpenAPI schema, then put that schema in the GPT editor's action section.

I'm not giving you a made-up Zambo Actions import link. No ready-to-import schema was verified in the Zambo pages checked for this guide. Native MCP is the shorter path where your account supports it.

Cursor

Create or edit either:

.cursor/mcp.json for this project
~/.cursor/mcp.json for all projects

Merge this:

{
  "mcpServers": {
    "zambo": {
      "url": "https://zambo.dev/mcp"
    }
  }
}

Cursor supports remote streamable HTTP natively. No npx bridge needed. Manage servers from the Customize page and allow the tools you want the agent to use.

Windsurf / Cascade

Check which agent you're using. The current docs distinguish legacy Cascade from the newer Devin local agent. They do not use the same config.

For Cascade, edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "zambo": {
      "serverUrl": "https://zambo.dev/mcp"
    }
  }
}

For the newer Devin local agent, use its CLI configuration instead:

devin mcp add -s user zambo https://zambo.dev/mcp

VS Code / Copilot

Open the command palette.
Run MCP: Add Server.
Choose an HTTP server.
Enter https://zambo.dev/mcp and name it zambo.
Choose global or workspace scope.
Review the trust prompt.
Open chat's tool picker and enable the Zambo tools you want the agent to use.

Or create .vscode/mcp.json:

{
  "servers": {
    "zambo": {
      "type": "http",
      "url": "https://zambo.dev/mcp"
    }
  }
}

Notice servers, not mcpServers. Copying the Cursor wrapper into this file is the wrong shape.

Claude Code

claude mcp add --transport http --scope user zambo https://zambo.dev/mcp

Use /mcp to inspect the server. Adding the config is not the same as a successful connection.

Gemini CLI

gemini mcp add --transport http --scope user zambo https://zambo.dev/mcp

Or merge this into ~/.gemini/settings.json:

{
  "mcpServers": {
    "zambo": {
      "httpUrl": "https://zambo.dev/mcp"
    }
  }
}

httpUrl is streamable HTTP. url selects the older SSE transport in Gemini CLI, so don't blindly copy Cursor's config field here.

Manus

Manus documents custom hosted MCP under Settings > Integrations > Custom MCP Servers > Add Server.

Enter zambo as the name and https://zambo.dev/mcp as the HTTPS endpoint, then test the connection and check that tools appear.

There is an auth caveat: the custom-server guide describes credential fields but does not explicitly confirm anonymous-server handling. Zambo's no-auth connection in Manus is unverified here. Don't supply fake credentials if the dialog requires them.

n8n

n8n's MCP Client node documents Server Transport, MCP Endpoint URL, authentication none, tool selection, and manual/JSON inputs. The exact streamable-HTTP menu option was not spelled out in the inspected page, so that part is unverified here. Use https://zambo.dev/mcp only if your node offers the matching streamable-HTTP transport; do not substitute an SSE endpoint.

The Honest Boundary

Not every assistant has a documented bring-your-own-MCP menu.

Dot is ambiguous. Neither inspected official page supplied a verified add-custom-MCP procedure. Zambo connection steps remain unverified.

Personal.ai: The inspected official material documents its own API and Zapier integration, not a custom MCP client setup. Add-Zambo is unverified.

Genspark: No official add-Zambo/custom-MCP client procedure was verified from the pages checked.

If your assistant isn't listed, look for documented remote MCP client support. A built-in connector catalog, browser access, or a product exporting its own MCP server does not establish that it can import any hosted server.

No fake universal setup instructions. If a client can't connect, use a documented compatible client or build an explicit API integration.

Connected? Do This First

Ask your AI:

Use the Zambo MCP server.
Call new_session with no arguments.
Save the returned session_id.
Use that exact value as _session_id on related Zambo calls.
Return the receipt link and the recorded status, not just "done".

Then give it your actual task.

zambo_universal is the entry point. Its need field is your job in plain language.

Switch AIs. Keep the Receipts.

Here's the habit that matters: keep the same _session_id for the same job.

Connect Zambo to the second AI with the same endpoint. Privately give it the session ID and tell it to inspect the trail before continuing.

Same job ID. Same Zambo trail. Different host AI.

This carries the context Zambo recorded. It does not automatically transfer every message, hidden model state, local file, or account permission from your old chat. The second AI still needs the tools and access required for the next step.

Change or omit the ID when you want to start clean.

And keep privacy in mind. A public receipt link is shareable evidence, not a place for credentials or confidential payloads.

That's the Whole Loop

Connect. Run. Inspect. Keep the link.

Switch the AI when you want. Carry the Zambo session trail instead of relying on the old model to narrate what happened.

Zambo is the hosted implementation. AER-1 is the open draft spec for the receipt format. Not an adopted IETF standard. Not proof that an AI can never be wrong.

Just a smaller gap between "the agent said it did" and a record you can actually check.

Try it now, no install: https://zambo.dev/demo/

Verify any receipt: https://zambo.dev/verify/

The spec: https://datatracker.ietf.org/doc/draft-zambo-aer1/

Your AI said "done." For $0.99, prove it: https://zambo.dev/day-pass/

DEV Community