Vibeleaderboard
← All Intel
Intel / post

Browser tools join MCP tools in codemode for agent scripting

Source
x.com
Date
dax@thdxr

that last point - all the browser tools are now available in codemode which means your agent can compose all your mcp tools along with browser actions eg open a url, fetch a list of things, iterate over them and call your mcp tool in a loop, then continue very powerful https://t.co/hfy71hnQ1p

Context

dax, who builds the open-source coding OpenCode, is drawing to one change in OpenCode Browse 0.1.0: an OpenCode agent can now drive a web browser from inside a short program it writes, alongside its other tools. His post quotes OpenCode's release announcement, which says Browse was rewritten in Rust, uses 25 times less memory, is 23 times smaller, no longer depends on Playwright, a browser-automation library, and exposes browser operations directly to the agent. "That last point" refers to that final item.

Browse is an server paired with a Chromium browser extension. MCP, the Model Context Protocol, is a standard way to hand an agent a set of tools; Browse's tools let it open background tabs in the user's real browser, with their logins and extensions, read pages, click, type and take screenshots. Code Mode is an OpenCode feature in which the agent writes a small JavaScript program through a tool named execute, and that program can call and combine any tool in its catalog. OpenCode's documentation says this helps with running independent calls in parallel and with processing results without adding every intermediate value to the model's , the text the model has to read at each step. Tools from MCP servers are reached this way by default.

Browse's README says OpenCode, because it composes tools in Code Mode, gets each browser operation as a separate tool, such as navigate, snapshot and click, while other MCP clients get a single execute tool that runs browser scripts inside Browse itself. Placing those browser tools in the same catalog as the user's other MCP tools is what dax describes: one program can open a page, pull a list of items from it, call another MCP tool for each item in a loop, and then carry on. Done as separate tool calls, each of those steps would be another round trip through the model. Browserbase has made a similar case for letting the model write a program that coordinates services.

dax calls this very powerful, but neither post reports measurements, and the size and memory figures are OpenCode's own. The README also states the risk that comes with the design: Browse drives the browser the user is signed in to, so it assumes that anything an agent reads, such as a web page or an email, may try to steer it.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
More from dax
Recommended reads
Comments

Checking sign-in…

Loading comments…