The latest articles on DEV Community by Japneet Kalkat (@jassskalkat). https://dev.to/jassskalkat https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3945024%2Fd3a97888-4d31-45cf-a0b9-00600afb5310.jpeg https://dev.to/jassskalkat en Japneet Kalkat Fri, 22 May 2026 01:31:07 +0000 https://dev.to/jassskalkat/i-built-agentyc-to-make-browser-automation-for-ai-agents-more-predictable-4pf5 https://dev.to/jassskalkat/i-built-agentyc-to-make-browser-automation-for-ai-agents-more-predictable-4pf5 <p>A lot of browser tools for AI agents look impressive in demos but become hard to trust in real work.</p> <p>They click around, extract something fuzzy, and when they fail, it is not obvious why.</p> <p>I built <a href="https://github.com/distillation-labs/agentyc" rel="noopener noreferrer">Agentyc</a> to make this simpler.</p> <p>Agentyc is an MCP-first browser runtime for coding agents. The goal is straightforward:</p> <ul> <li>give the agent a real browser</li> <li>make the actions deterministic</li> <li>make the page state inspectable</li> <li>avoid hidden LLM fallback for common extraction tasks</li> </ul> <h2> Why I built it </h2> <p>I wanted something that felt closer to debuggable browser automation than a black-box browser agent.</p> <p>If an agent opens a page, clicks a button, types into a form, or extracts a table, I want to know:</p> <ul> <li>what page state it saw</li> <li>which element it targeted</li> <li>what changed after the action</li> <li>what the console and network looked like if it broke</li> </ul> <p>That is the problem Agentyc tries to solve.</p> <h2> What it can do </h2> <p>With the default MCP server, a coding agent can:</p> <ul> <li>navigate pages</li> <li>click, type, scroll, and upload files</li> <li>inspect page state with stable element refs</li> <li>take screenshots</li> <li>read HTML</li> <li>inspect cookies</li> <li>inspect console and network logs</li> <li>extract links, tables, lists, form fields, and key-value blocks in a deterministic way</li> </ul> <h2> Quick start </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>uv tool <span class="nb">install </span>agentyc agentyc init agentyc </code></pre> </div> <p><code>agentyc init</code> writes a guide file your coding agent can read.</p> <p><code>agentyc</code> starts the MCP server.</p> <h2> A simple example </h2> <p>One of the most useful parts is <code>browser_get_state</code>.</p> <p>Instead of dumping a whole page every time, the agent can ask for a compact state view, get stable refs like <code>e123</code>, and act on those refs.</p> <p>That makes loops like this possible:</p> <ol> <li>open a page</li> <li>get compact state</li> <li>click or type using stable refs</li> <li>poll with <code>since_hash</code> until the page changes</li> <li>extract only the thing you need</li> </ol> <p>That ends up being cheaper in tokens and easier to debug.</p> <h2> Why deterministic matters </h2> <p>A lot of browser + AI workflows break because the tool quietly switches from structured automation to fuzzy interpretation.</p> <p>Agentyc takes the opposite approach for common tasks like extracting a table, list, links, or form fields:</p> <ul> <li>if there is a deterministic route, use it</li> <li>if there is not, return a clear error</li> </ul> <p>That tradeoff is less magical, but much easier to trust.</p> <h2> Shared browser mode </h2> <p>Another feature I wanted was the ability to attach to an existing Chrome or Chromium session.</p> <p>That means a human and an agent can work in the same browser more naturally:</p> <ul> <li>launch a browser with remote debugging</li> <li>attach Agentyc to it</li> <li>let the agent work in a tab or window</li> <li>keep visibility into what is happening</li> </ul> <h2> Who this is for </h2> <p>Agentyc is useful if you are building:</p> <ul> <li>coding agents that need browser access</li> <li>MCP tools for web workflows</li> <li>automation that needs better debugging than a black-box agent gives you</li> <li>browser workflows where predictable behavior matters more than magic</li> </ul> <h2> What I want feedback on </h2> <p>Agentyc is open source and still early.</p> <p>I would love feedback on:</p> <ul> <li>the MCP tool surface</li> <li>the state model</li> <li>shared-browser workflows</li> <li>places where the browser loop still feels too heavy</li> </ul> <p>GitHub: <a href="https://github.com/distillation-labs/agentyc" rel="noopener noreferrer">https://github.com/distillation-labs/agentyc</a></p> ai programming python opensource