Product
Start the real host.
See what it called.
We start Claude Code, Cursor, Codex, and the other hosts we run, send the requests a person would type, and keep the version and the tool it called.
The URL is the start. The host is the test.
Paste the server. We check which hosts can reach you, which tools did the job, which step died, and what changed after you shipped.
Connect the MCP
Just the URL. No SDK needed.
Paste the server URL
URL and credentials. That is the first artifact. Named hosts begin connecting.
anon · 3
authed · 7
No SDK and no code change
You do not instrument the host or wrap the server. Auth rides with the URL.
Turn discovery into checks
Listed tools and workflows become explicit jobs. Discovery is the start of the record, not the product.
Host check
Know which hosts can actually reach you.
CursorREACHED
ClaudeREACHED
ChatGPTNO ANSWERCheck the hosts people use
Cursor, Claude, VS Code, ChatGPT. The clients that call you. Not a generic ping.
Catch more than downtime
Initialize, the catalog, and the call can still fail when the URL is up.
Same server. Different hosts.
We record whether each host can connect, list tools, and complete a call.
Tools and jobs
Prove which tools did the job.
Cursor20 TOOLSClaude20 TOOLS
VS Code20 TOOLSSee the catalog each host got
What that client was actually shown. Listed is not called.
Prove the call completed
Whether the tool ran, not only whether it appeared.
Output matches what the tool description claims?
Prove the job was done
The defined outcome after the call. Called is not the work.
Multi-step
Catch the step that killed the path.
Walk the path
Search, then fetch, then write. The first call, then the call that needs the prior result.
↓ extract
Stop at the failed step
That hop is the finding. Not a vague chain error.
ClaudeDIED AT FETCHCursorDIED AT WRITEChatGPTFINISHEDCompare the same path across hosts
Cursor finished. Claude died on fetch. Same server.
What changed
Pin what broke after you shipped.
Compare before and after the release
The tested response, then the one that shipped.
Pin the first time it broke
Timestamp, run, and host. Not a guess.
Separate drift from a working call
Catalog moved. The job may still pass. You get both.
FAQ
Common questions
| Question | Answer |
|---|---|
| Do I need an SDK? | No. Paste the MCP URL and auth. No SDK and no code change. Live Usage starts the host from there. |
| We ping the URL and it's up. Why isn't that enough? | A 200 on the address does not mean Claude Code called the tool, or that Cursor did the same thing Codex did. Uptime checks the URL. A sitting shows the host, the version, and the tool it called. |
| Which hosts do you start? | Host Compatibility starts Claude Code, Cursor, Codex, Muse Code, Goose, Cline, OpenCode, and Devin. Each sitting keeps the version, the request, the tool that host called, and the server's answer. Copilot, Mistral, Qwen Code, Antigravity, Kiro, and Grok are listed and not started. |
| Discovery passed. Aren't we done? | Listed is not called. Called is not the job. A tool can sit in the catalog and fail every time a host tries to use it. |
| Do you test multi-step paths? | Yes. Search, then fetch, then write. We walk that chain on the same MCP and stop where it breaks, per host. |
| What changed after we shipped? | We compare catalogs, initialize results, and call verdicts across hosts, runs, and releases. You get the first time it broke, not a guess. |
| What do we actually get back? | Which hosts connected, which tools did the job, which step died, and what changed after you shipped. Reviewable record, not a status ping. |
Your biggest blind spots sit outside your logs,inside hosts and paths you don't control.
We start the host, send the request a person would type, and keep the version and the tool it called.
Cover your blind spots