Your MCP server works.
But will an agent choose the right tool?
Run realistic tool-selection tests, uncover ambiguous tools, and improve your MCP server before agents fail in production.
No signup. Your tools are never executed.
A six-tool CRM server, deliberately ambiguous: get_customer, search_customers and search_contacts absorb each other's traffic. See the confusion matrix and the simulated fix.
Why valid schemas still fail
get_customer vs search_customers
Both descriptions say they retrieve customer information. Neither says whether the identifier must already be known. A linter passes both; an agent flips a coin.
required: ["id"] — but which id?
The schema is valid. The model still passes an email into id because nothing documents the cus_ prefix.
cancel_invoice — no warning
Destructive tools that don't state their consequences get called casually. That's not a schema error. It's a behavioral one.
Paste, measure, fix
- 1
Paste tools/list JSON
No signup, no connection required. The audit runs on your definitions alone — your tools are never executed.
- 2
Test selection behavior
Tool-selection tests run with structured output at temperature 0 — against a live model when configured, or the deterministic synthetic simulator (always labeled) when not. Demo audits score against human-authored fixtures.
- 3
Rerun before you trust a rewrite
Rewrites follow a strict shape — action, context, when to use, when not to — and the identical suite is rerun against them. Improvements are only claimed when the second evaluation supports them, in the same labeled mode.
| expected | selected instead | count |
|---|---|---|
| get_customer | search_customers | 4 |
| search_contacts | list_contacts | 2 |
| create_invoice | update_invoice | 2 |
Every misroute is explained: likely cause, then a concrete fix.
“Gets customer data.”
“Retrieve one customer by their unique customer ID. Use this tool when the customer ID is already known. Do not use it to search by name, email, or company.”
68/100 → 84/100 · identical test suite, rerun — regressions are shown too
Remote MCP endpoint connection ships next — with SSRF protection, response limits, and no tool invocation.
Deterministic rules and LLM judgments never mix. Static findings are reproducible byte-for-byte; behavioral tests are ground-truth validated and run at temperature 0; the final score is a fixed weighted formula that no model can invent. Read the full methodology →
Free Beta
AVAILABLE NOW- · Up to 20 tools per audit
- · 10 scored behavioral tests per audit
- · Synthetic evaluation, unlimited within fair-use limits
- · Live-model audits within daily beta limits
- · Shareable reports (7-day retention for anonymous audits)
- · JSON + Markdown export
No account required. No credit card. Beta limits exist to keep the service free and responsive.
Developer
COMING SOON- · Saved audit history
- · Higher limits & larger collections
- · Multiple model presets
- · Scheduled re-audits
Pricing not final — nothing is purchasable yet.
Team
COMING SOON- · Team workspace
- · CI / GitHub integration
- · Regression alerts
- · API access
Pricing not final — nothing is purchasable yet.
Know whether agents choose the right tool.
Test your MCP tools against realistic user requests. Find ambiguous tools, missing descriptions, parameter confusion, and selection failures before your users do.
Run a free audit