Skip to content
fastbrowsefastbrowse

Coding agents and MCP

Add a browser skill to Claude Code and Codex

Install the fastbrowse skill so Claude Code or Codex can browse a site and come back with an answer, its quotes and a run status.

·all posts

How the skill picks a browserA coding agent calls fastbrowse through the skill. The skill uses an open Chrome connection first, then a local Chrome, then a cloud browser, and returns an answer, quotes and a status.ANSWER, QUOTES, STATUSCoding agentClaude Code, Codexfastbrowseskill: CLI or MCP1. Chrome already opensigned in, no credential2. Local Chromemodel key only3. Cloud browserwhen asked, or no local
The skill tells your coding agent when to call fastbrowse and which browser to use: one already open, then a local Chrome, then a cloud browser.

The fastbrowse skill lets a coding agent hand a browser task to fastbrowse and get back an answer, the quotes behind it and a run status it can branch on.

The skill is a set of instructions. It tells the agent when to use fastbrowse, which command or tool to call, which browser to prefer, and how to read a result that is not complete. You still configure the model key, and a browser key if the agent may use a cloud browser.

Install it

The Skills CLI installs it from GitHub:

npx skills add agent-labs-dev/fastbrowse --skill fastbrowse

Read the skill before you install it, since it decides when your agent acts on a website.

To install by hand, copy skills/fastbrowse to .claude/skills/fastbrowse for Claude Code or .agents/skills/fastbrowse for Codex, keeping its references directory beside SKILL.md. Start a new session afterwards, because skills load when a session starts.

Use it

The agent picks fastbrowse up from the skill's description, or you name it. In Claude Code: /fastbrowse Find the top story on Hacker News and return its title with a citation.

The skill tries browsers in this order:

  1. An existing Chrome connection, if one is open.
  2. A local Chrome.
  3. A cloud browser, when you ask for one or local browsing is unavailable.

A local browser needs only the model key, and an already signed-in browser needs no credential.

Behind the skill, the agent calls the CLI or an MCP server. Either way the run reports a status. complete means the task was verified, and the other statuses say what was missing: an input, an authorization, evidence for a claim, or access to the page.

Limits and writes

The skill does not authorize writes. For a form, give the agent the URL and the values, and say whether to stop before submitting or to submit. Without that, an irreversible action stops at the confirmation step.

The skill sets no spending limits. Pass limits yourself, or put a budget in the task. A cloud browser's charge is added when it stops, separately from any cap on model spend.

Use fastbrowse with your agent is the install reference. MCP browser server with spending limits covers the same tool with its limits fixed on a server.

Next: your first run.