Know when AI agents stop calling your MCP tools.
Daily checks in Claude, Cursor, ChatGPT and Codex. When your tool stops getting called, you know, and you know why.
Add your server URL when you book and get a free call-rate report within 48 hours.
Who calls track_order
MCP analytics · acme-shop-mcp · last 14 days · demo data
| Agent | Share of runs, 0 to 100% | Now | Why |
|---|---|---|---|
| ClaudeOpus 5.5 | 31% | CauseClaude app update | |
| ChatGPTGPT-5.5 | 34% | CauseModel update | |
| CursorSonnet 5.5 | 38% | CauseYour release v2.4 | |
| CodexGPT-5.5-Codex | 76% | Steady | |
| Gemini3.1 Pro | 77% | Steady |
You only see the calls agents make. Never the ones they skip.
No error. No log line. Your server just gets quieter.
Watch both screens: the agent looks through what it has, finds nothing, and asks the customer which shop it was. Your tool, in red, was right there.
How it works
Paste your MCP server URL.
We read its tool list and build a stand-in. Your server isn't called unless you opt in.
Add the prompts your users type.
Every day we run them in each agent: Claude, Cursor, ChatGPT, Codex and Gemini.
See the call rate per agent.
When it drops, you get an alert with the cause (your release, the model, or the agent’s update) and a suggested fix.
Free while we build it with the first teams.
Pro from $149 a month at launch. Reserve the early-access price now; nothing to pay today.
Questions
Does Uncalled call my MCP server?
Not by default. It reads your tool list once and builds a stand-in from it, so no test traffic reaches your server. Live checks that call your server with a test key and verify the response are planned as an opt-in.
Which agents does it check?
Claude, Cursor, ChatGPT, Codex and Gemini at launch. Each one is run with the prompts your users type, every day.
What makes an agent stop calling a tool?
Three things, mostly: your own release changed the tool description or schema; a new model version weighs tools differently; or the agent app updated its built-in tools or prompts. Uncalled checks all three and names the one that matches.
How is this different from logging tool calls, like PostHog MCP Analytics?
Logs show the calls that reached your server. Uncalled shows the ones that never happened, per agent, with the cause.
How is this different from writing evals?
Evals test your tool in the setup you wrote them for, on the day you run them. Uncalled reruns your prompts every day in each agent, so you see when a new model or an agent update changes what gets called.
What is in the free report?
The prompts your users would type, run repeatedly in real agents against a stand-in of your tools. You get how often each agent called each tool, and what it did instead when it did not. It arrives by email within 48 hours.
Who is behind it, and what does it cost?
One founder, building it with the first teams. The daily checks are not live yet, so reports are run by hand for now. Free during early access; Pro is planned from $149 a month at launch. Write to hello@uncalled.app.