fastbrowse-mcp serves one browse tool over MCP. Its limits are set when the server starts. A call can ask for less, never more.
Flags the server fixes
A calling model cannot override these:
| Flag | Effect |
|---|---|
--local, --headed, --profile DIR, --downloads DIR, --cursor, --no-cursor | as for the CLI, fixed for every call |
--cloud-profile ID | every call runs as that cloud profile |
--allow-authorize | lets a call pass authorize; without it an irreversible action always stops at needs_confirmation |
--secret NAME=ENV_VAR@ORIGIN | typed when a call's start page is on ORIGIN; the model sees NAME only |
--bitwarden ITEM | a vault login a call may name |
--max-steps N, --max-dollars N, --max-seconds N | ceilings per call, unlimited by default |
--max-concurrent N | runs at once, default 1; further calls wait |
With --allow-authorize off, every irreversible action stops for confirmation. With the three ceilings set, you know the most a call can cost before the first one runs.
Stdio or HTTP
Over stdio, the server runs for as long as the client that launched it. With --transport http it serves streamable HTTP at /mcp. Every request except /healthz then needs the bearer token in FASTBROWSE_MCP_TOKEN, and the server will not bind beyond loopback without one.
Add it to a client
claude mcp add fastbrowse -e TYPESAFE_API_KEY=... -e OPENROUTER_API_KEY=... -e BROWSER_USE_API_KEY=... \
-- uvx --from fastbrowse fastbrowse-mcp --max-dollars 0.25
A client configured by JSON, such as Claude Desktop, uses the same command, uvx --from fastbrowse fastbrowse-mcp, with those keys in its env. A run takes seconds to minutes, so raise the client's tool timeout. In Claude Code that setting is MCP_TOOL_TIMEOUT.
What the tool returns
The tool returns the answer with numbered citation links, typed fields if the call asked for them, the quotes behind those fields, the run status and a next step. It reports progress at every step. The status tells the model whether the answer was verified.
Use it from an MCP client is the server reference. Add a browser skill to Claude Code and Codex covers the skill route, and the safety model explains what the authorization gate judges.
Next: your first run.