Bookbag tools for any MCP-native agent.
Claude Desktop, Cursor, custom MCP clients — drop in this server and they call Bookbag's gate + eval as first-class tools. stdio or streamable-http. MIT licensed.
What it ships
10 MCP tools
Gate surface: start_run, tool_call, output, end_run, await_decision, fetch_policy_bundle. Eval surface: list/import/run/get eval suites.
Two transports
stdio for local spawn (Claude Desktop). streamable-http for remote clients (Cursor, CI, custom agents).
Zero-dep server
Node 18+ builtins only. Small Docker image. Health endpoint + log redaction built-in.
CLI binary
Install once globally (npm i -g bookbag-mcp-server) — bookbag-mcp start / --help / --transport http.
Three ways to connect
Local spawn for Claude Desktop. Remote HTTP for Cursor. Direct JSON-RPC for testing.
// ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"bookbag": {
"command": "npx",
"args": ["-y", "bookbag-mcp-server"],
"env": {
"BOOKBAG_API_KEY": "bk_key_...",
"BOOKBAG_BASE_URL": "https://app.bookbag.ai"
}
}
}
}
// Restart Claude Desktop. Ten new tools (bookbag_start_run, bookbag_tool_call, etc.)
// are available. Claude can gate its own tool calls through Bookbag.MCP Server FAQs
Frequently Asked Questions
Point Claude or Cursor at Bookbag. Ten tools show up.
Join the teams shipping safer AI with real-time evaluation, audit trails, and continuous improvement.