DEV Community

Japneet Kalkat
Japneet Kalkat

Posted on

I built Agentyc to make browser automation for AI agents more predictable

#ai #programming #opensource #python

A lot of browser tools for AI agents look impressive in demos but become hard to trust in real work.

They click around, extract something fuzzy, and when they fail, it is not obvious why.

I built Agentyc to make this simpler.

Agentyc is an MCP-first browser runtime for coding agents. The goal is straightforward:

  • give the agent a real browser
  • make the actions deterministic
  • make the page state inspectable
  • avoid hidden LLM fallback for common extraction tasks

Why I built it

I wanted something that felt closer to debuggable browser automation than a black-box browser agent.

If an agent opens a page, clicks a button, types into a form, or extracts a table, I want to know:

  • what page state it saw
  • which element it targeted
  • what changed after the action
  • what the console and network looked like if it broke

That is the problem Agentyc tries to solve.

What it can do

With the default MCP server, a coding agent can:

  • navigate pages
  • click, type, scroll, and upload files
  • inspect page state with stable element refs
  • take screenshots
  • read HTML
  • inspect cookies
  • inspect console and network logs
  • extract links, tables, lists, form fields, and key-value blocks in a deterministic way

Quick start 

uv tool install agentyc
agentyc init
agentyc
Enter fullscreen mode Exit fullscreen mode

agentyc init writes a guide file your coding agent can read.

agentyc starts the MCP server.

A simple example

One of the most useful parts is browser_get_state.

Instead of dumping a whole page every time, the agent can ask for a compact state view, get stable refs like e123, and act on those refs.

That makes loops like this possible:

  1. open a page
  2. get compact state
  3. click or type using stable refs
  4. poll with since_hash until the page changes
  5. extract only the thing you need

That ends up being cheaper in tokens and easier to debug.

Why deterministic matters

A lot of browser + AI workflows break because the tool quietly switches from structured automation to fuzzy interpretation.

Agentyc takes the opposite approach for common tasks like extracting a table, list, links, or form fields:

  • if there is a deterministic route, use it
  • if there is not, return a clear error

That tradeoff is less magical, but much easier to trust.

Shared browser mode

Another feature I wanted was the ability to attach to an existing Chrome or Chromium session.

That means a human and an agent can work in the same browser more naturally:

  • launch a browser with remote debugging
  • attach Agentyc to it
  • let the agent work in a tab or window
  • keep visibility into what is happening

Who this is for

Agentyc is useful if you are building:

  • coding agents that need browser access
  • MCP tools for web workflows
  • automation that needs better debugging than a black-box agent gives you
  • browser workflows where predictable behavior matters more than magic

What I want feedback on

Agentyc is open source and still early.

I would love feedback on:

  • the MCP tool surface
  • the state model
  • shared-browser workflows
  • places where the browser loop still feels too heavy

GitHub: https://github.com/distillation-labs/agentyc

Top comments (0)