forecall-cli MCP server

Forecall CLI

The open-source parts of Forecall: the linter that scores how well an MCP server's tools are described for the AI agents that read them, and the forecall command that runs it on your machine and connects your AI clients to the Forecall Failure KB.

npx forecall lint tools.json

Forecall lint · tools.json
Scoring rules v2

Average score                42.5 / 100
Tools                        2
Near-identical descriptions  0

Across the server
  No server-wide issues.

Tools

  6 / 100  search
    Purpose 6/20 · When to use 0/20 · Arguments 0/25 · Return value 0/15 · Constraints and side effects 0/10 · Examples 0/10
    Major     restates_name         The description only restates the name (2 words).
    Major     too_short             The description is too short (2 words).
    Major     no_usage_context      Says nothing about when to use it, or when not to.
    Major     param_no_description  Arguments without a description: q.
    Minor     loose_string_param    String arguments with no constraint or example: q.
    Minor     no_required           No `required` list, so every argument is optional.
    Minor     no_return_info        Says nothing about what it returns.
    Minor     no_constraints        No constraints, side effects, auth or rate limits described, and no annotations.

  79 / 100  get_forecast
    Purpose 17/20 · When to use 20/20 · Arguments 18/25 · Return value 7/15 · Constraints and side effects 7/10 · Examples 10/10
    No issues.

{
  "tools": [
    {
      "name": "search",
      "description": "Search records.",
      "inputSchema": { "type": "object", "properties": { "q": { "type": "string" } } }
    },
    {
      "name": "get_forecast",
      "description": "Returns the weather forecast for a city for the next 7 days, with the high and low temperature in Celsius for each day. Use it when the user asks about upcoming weather; for the current conditions, use get_current_weather instead. Example: city \"Tokyo\", days 3.",
      "inputSchema": {
        "type": "object",
        "properties": {
          "city": { "type": "string", "description": "The city's name, such as Tokyo." },
          "days": { "type": "integer", "minimum": 1, "maximum": 7, "description": "How many days to return, 1 to 7." }
        },
        "required": ["city"]
      },
      "annotations": { "readOnlyHint": true }
    }
  ]
}

Each tool is scored out of 100 on six parts, and the server as a whole is checked for too many tools, near-identical descriptions and shared descriptions. Nothing leaves your machine. --json prints the report as JSON. --fail-under <n> fails a CI job when the average is below n, --min-tool-score <n> when any tool is, and --fail-on <severity> on any issue that severe. Get the JSON from your own server with npx forecall dump, or paste it at forecall.dev/en/lint for a page you can share.

How to use them: forecall.dev/en/docs/cli and each package's README. The Forecall service itself (the website, the dashboard, the API and the Failure KB's server) is not in this repository.

GitHub Action

Score your tools on every pull request, with the score and the findings in the job summary:

- uses: forecall/forecall-cli@v1
  with:
    file: tools.json          # or url: https://… (a public Streamable HTTP server)
                              # or command: node build/index.js (your server over stdio)
    min-tool-score: 40        # optional gates, as in the CLI
    fail-on: critical

Give exactly one of file, url and command. The outputs are score (the average), tools and report (the JSON report's path). It needs Node 22 or later on the runner, sends nothing to Forecall and needs no key.

Development

The versions of Node.js and pnpm are in mise.toml.

mise install
pnpm install
pnpm check       # Biome, the type checks and the tests of both packages and the action

• pnpm format formats and fixes what Biome can.
• pnpm --filter forecall build writes the command to packages/cli/dist/forecall.js; run it with node packages/cli/dist/forecall.js lint <file>.
• node packages/lint/pack.ts <dir> and node packages/cli/pack.ts <dir> make the tarballs npm would get, and scripts/try-package.sh <lint|cli> <tarball> installs one and runs it.
• LINT_VERSION in packages/lint/lint.ts is the version of the scoring rules. Any change that can give some input a different score, component, issue code or severity raises it, and updates the expected results in packages/lint/fixtures/cases.json.

Releases

A pull request that raises a package's version starts the Release workflow when it lands on main. The workflow stages every version npm does not have yet (@forecall/lint first) with provenance, and each goes live only when a maintainer approves it on npm with 2FA. Maintainers can also run the workflow by hand on main and pick one package.

The action is released apart from the packages: a maintainer publishes a GitHub Release v1.x.y to the Marketplace and moves the v1 tag to it. A release that raises the action's pinned forecall version says so, since a new rules version can move scores past a gate.

Contributing and security

See CONTRIBUTING.md. Report vulnerabilities privately, as SECURITY.md says, not in a public issue.

License

Apache-2.0. Copyright 2026 Mirai Studio, Inc. See LICENSE.