Forecall

Lint your MCP tool descriptions before AI agents call them

4 followers

Lint your MCP tool descriptions before AI agents call them

4 followers

AI agents choose and call MCP tools by reading their names, descriptions and input schemas. Vague text means the wrong tool, wrong arguments, or no call at all. Forecall scores each tool out of 100 on six items (purpose, when to use, arguments, return value, constraints, examples), finds tools that are easy to mix up, and explains every issue with a code and severity. Free on the web with no sign-up, or run npx forecall lint locally and in CI.

Free

Launch tags:API•Developer Tools•Artificial Intelligence

Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building

Promoted

Maker

📌

Hi Product Hunt! AI agents pick a tool by reading its name, description and input schema, and nothing else. When we scored 8 well-known MCP servers (126 tools), their averages ranged from 33 to 70 out of 100, and Notion's official server alone had 57 pairs of tools an AI can easily mix up. Forecall is a static linter for that text. It scores each tool on six items, explains every issue with a code and a severity, and points out tools that are easy to confuse. It's free on forecall.dev with no sign-up, and the same rules run locally with npx forecall lint (nothing is uploaded), so you can fail a CI job when the average drops. Next, we're building Evaluations that measure how often real models choose and call your tools correctly. I'd love your feedback on the scoring rules: what should a good tool description say?

Report

1mo ago

Maker

A bit more on what the scores look like in practice 👇 It isn't about how many tools you have. The Playwright MCP server has 25 tools and zero confusable pairs (average 41.9), while Notion's official server has 24 tools and 57 pairs (average 33.0). Notion's lowest-scoring tools are API-query-data-source (20/100), API-create-a-comment (24) and API-update-a-data-source (24). Three edits worth trying first, each one of the six items: 1. A "when to use" line that names the neighbor: "Use this to …; to … instead, use X." 2. One sentence on what it returns. 3. One example call with real argument values. To try it, paste the JSON your server returns for tools/list at forecall.dev. No sign-up needed. Post your average below and I'll reply with the first change I'd make 🙌

Report

1mo ago

Maker

Weekly update from the maker:

Evaluations are live: real models (Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra and 6.1 Sol) answer utterances written for each of your tools, and you see tool choice, argument validity and false calls per model. On the Pro and Team plans.
The Failure KB is live: an MCP server of known MCP tool failures and their workarounds, in the official MCP Registry. Connect your agent with npx forecall setup.
The linter is now open source (Apache-2.0): https://github.com/forecall/forecall-cli. Two CI gates a reader asked for shipped the same day: --min-tool-score and --fail-on.
No tools/list JSON at hand? See Notion's official server, scored: https://forecall.dev/en/lint/sample
New data: in Smithery's 46 most-used servers, 69% of 1,118 tools never say when to use them. https://dev.to/nomunomu0504/we-checked-46-popular-mcp-servers-on-smithery-7-in-10-tools-never-say-when-to-use-them-4gb9

Report

30d ago

Pros

Cons

Reviews

Weekly update from the maker:

Evaluations are live: real models (Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra and 6.1 Sol) answer utterances written for each of your tools, and you see tool choice, argument validity and false calls per model. On the Pro and Team plans.
The Failure KB is live: an MCP server of known MCP tool failures and their workarounds, in the official MCP Registry. Connect your agent with npx forecall setup.
The linter is now open source (Apache-2.0): https://github.com/forecall/forecall-cli. Two CI gates a reader asked for shipped the same day: --min-tool-score and --fail-on.
No tools/list JSON at hand? See Notion's official server, scored: https://forecall.dev/en/lint/sample
New data: in Smithery's 46 most-used servers, 69% of 1,118 tools never say when to use them. https://dev.to/nomunomu0504/we-checked-46-popular-mcp-servers-on-smithery-7-in-10-tools-never-say-when-to-use-them-4gb9