Getting started
What Forecall measures, what you can do for free, including the Failure KB for your agents, what the paid plans add, and what static scoring cannot tell.
On this page
An AI agent chooses a tool by reading its name, its description and its input schema. When the text leaves out what the agent needs, it has less to go on: it may pick the wrong tool, fill in the wrong arguments, or skip a call it should make. Forecall measures how well your MCP tools are written for that reader, and shows you what to fix.
What Forecall measures
Forecall reads what your MCP server returns for tools/list and scores each tool out of 100 on
six things a model needs to choose and call it:
- its purpose
- when to use it, and when not to
- its arguments
- its return value
- its constraints and side effects
- examples
It also looks across the whole server: too many tools, tools whose descriptions are nearly the same, and tools that share the same description. Each finding comes with an issue code and a severity (critical, major or minor), so you can fix the worst first. Read your scores goes through the items and every issue code.
This is static scoring: Forecall reads the text and never runs your server or a model. Measuring how often real models choose and call your tools correctly is the job of evaluations.
Free and paid
Free, with no sign-up:
- The linter on the web: paste your
tools/listJSON at the linter. You get a result page whose URL you can share. Anyone with the URL can see it, search engines do not list it, and it is deleted after 30 days. - The CLI:
npx forecall lint tools.jsonscores a file on your own machine with the same rules and sends nothing anywhere. Use it in CI to keep scores from dropping.
Free, with an account at app.forecall.dev:
- Servers and history: register your servers and follow their scores over time.
- API keys: score from your own scripts and CI through the REST API, within a small monthly allowance.
- The Failure KB: connect your AI agent with
npx forecall setup, so that it looks up known failures of MCP tools and their workarounds before and after a call, within a monthly allowance of lookups.
Paid, on the Pro and Team plans:
- Evaluations: real models choose and call your tools, and Forecall measures how often they get it right.
- Failures reported on your servers: what agents reported on your servers' tools, by error class, tool, model and client.
- Larger allowances for API keys and Failure KB lookups.
Plans and billing compares the plans and their prices.
What static scoring cannot tell
The score reads only the text of each description and schema, so keep these in mind:
- Short tools whose names already say what they do, such as
browser_close, score low even when models use them well. - A high score does not prove that models choose the tool well, and a low score does not prove that they choose it badly. The score says what the text gives them; a low score points to the text worth improving. In our measurements (2026-10-09, on chrome-devtools-mcp and the reference filesystem and memory servers), current models chose the right tool in one step even from one-line descriptions and from pairs with near-identical descriptions, as long as the names and schemas differed.
- What the text cannot show is what happens over several steps (what a call returns, where its identifiers come from) and whether a sentence you add helps or misleads: in the same measurements, a longer description made one model skip a call it should have made and another pick the wrong tool. An evaluation measures that; the score does not.
- Scores are comparable only within the same version of the scoring rules. Every result shows the version it was scored with.
Try it
To see what a result shows before you have any JSON, open the sample result: Notion's official MCP server, scored with the current rules.
Then paste the JSON your server returns for tools/list at the linter, or let the CLI
get it from your server and score it on your machine. Replace <start command> with the command
that starts your server. Get your tools/list shows the other ways to get the
JSON.
npx forecall dump -o tools.json -- <start command>
npx forecall lint tools.json