tbro-mcp

Русская версия

An MCP server that gives Claude two web tools — web_search and web_fetch — that return plain text instead of HTML. Pages are rendered through the tbro terminal browser. The point is to spend fewer tokens on web lookups.

Claude already has web tools, so why this one? Two reasons. tbro reduces a page to just its readable text — no markup, scripts, navigation, or ads — and the result is plain enough that you can grep or trim it before it reaches the model. For a search that works out to roughly 8x fewer tokens.

The trade-off is JavaScript: there isn't any. Single-page apps and dynamic dashboards come back mostly empty, so the tool descriptions tell Claude to fall back to its built-in fetch for those. Static pages, docs, articles, and plain searches are where this earns its keep.

Requirements

• tbro on your PATH (or its path in TBRO_BIN). Install it from the tbro repo with cargo install --locked --path crates/app — it lands in ~/.cargo/bin.

• Node 18 or newer.

Build

npm install
npm run build

Add it to Claude Code

claude mcp add tbro -e TBRO_BIN=$HOME/.cargo/bin/tbro -- node /absolute/path/to/tbro-mcp/build/index.js

Restart Claude Code and the two tools appear. To wire it up by hand instead, add this to .mcp.json (project) or ~/.claude.json:

{
  "mcpServers": {
    "tbro": {
      "command": "node",
      "args": ["/absolute/path/to/tbro-mcp/build/index.js"],
      "env": { "TBRO_BIN": "/Users/you/.cargo/bin/tbro" }
    }
  }
}

Tools

• web_search(query, max_chars?) — run a search, get the results as text.

• web_fetch(url, max_chars?) — pull a single page as text.

max_chars caps the reply (default 8000). TBRO_BIN points at the binary if it isn't on your PATH.

Notes

Arguments are passed to tbro via execFile, not a shell, so there's nothing to inject. Each call times out at 30 seconds. test/smoke.mjs starts the server over stdio and calls both tools — run it with node test/smoke.mjs after a build.

License

MIT.