WORKS SAAS
Forecall
Measures whether AI agents choose and call your MCP tools correctly, and shows what to fix. The tool-description linter is open source.
Overview
How an agent calls an MCP tool depends on how the tool is described. Forecall starts with a linter that scores tool descriptions from 0 to 100, then measures whether agents actually choose and call the tools correctly, and shows what to fix.
Pieces
- forecall-cli:
npx forecall lint tools.jsonscores each tool on purpose, when to use it, arguments, return value, constraints and examples. Open source, with a GitHub Action - forecall-mcp: a small relay that connects stdio-only clients such as Claude Desktop to the Forecall Failure KB, an MCP server where agents look up known failures and workarounds before and after calling a tool
- forecall.dev: the measurement service and dashboard
Technical notes
- Scoring rules are versioned and separate server-wide issues (near-identical descriptions) from per-tool ones
- The KB is a Streamable HTTP MCP server; the relay exists only for clients that cannot reach a remote server