Contracts agents receive
Catch omitted schemas, vague names, credential-shaped arguments, unclear destructive actions, and missing output contracts on the live wire surface.
Open-source MCP readiness check
Your MCP can connect — and still fail in a real agent run.
Inspect the tool contracts models actually receive, find agent-facing defects, and record real execution evidence across selected model families.
npx --yes @coefficient-work/mcp-doctor@0.4.7 listInspect stays local and needs no model key. Optional evals are BYOK and send the task, schemas, calls, and results to the provider you select.
Live evidence / 0.4.7connected
$ mcp-doctor list
relayforge
stdio · 10 live tools · 52ms
Readiness
84/100
Grade B
Input contracts+ valid
Credential-shaped args+ none
Output schemas! 10 missing
Same task / three model familiescalls · results
OpenAIgpt-5.6-solproven · 1/1
Anthropicclaude-sonnet-5proven · 1/1
Googlegemini-3.7-flashproven · 1/1
Real non-error MCP result observedexecution proven
Surface
Live MCP
Artifact
Local Markdown
Model route
BYOK · explicit
Connection is the start
Test the part your users actually depend on.
Protocol success tells you a server answered. MCP Doctor examines whether the resulting tool surface gives agents a fair chance of succeeding.
Real tool execution
Run one task through selected models. A pass requires a real, non-error MCP result—not persuasive prose or provider finish metadata.
Evidence you can review
Keep the ordered calls, results, friction, token use, and named fixes in a local Markdown report that can travel with a release.
One release workflow
Discover. Inspect. Exercise. Review.
Start with the server configuration you already use. Add model traffic only when you want execution evidence.
Open the complete sample report →- 01No model key
Find the configured server
npx --yes @coefficient-work/mcp-doctor@0.4.7 list - 02Local report
Inspect the live contracts
npx --yes @coefficient-work/mcp-doctor@0.4.7 inspect <name> -o report.md - 03Optional · BYOK
Exercise a named task
Run the same task across explicit OpenRouter or direct-provider model selections. - 04Human review
Make the release decision
Read named fixes and the ordered replay. Commit the Markdown artifact if it belongs in your review process.
A deliberately narrow claim
Evidence for a decision. Not a certification badge.
+ MCP execution proven
At least one real, non-error tool result was observed.
— Not proven
Universal semantic correctness, security, load readiness, or every client.
+ Grade measures
Observable schema and agent-readiness signals.
— Grade does not measure
Whether your whole product is safe or production-ready.
Fits after your build path
Generated, framework-based, or hand-written.
MCP Doctor does not replace the tool you build or debug with. It examines the live surface after those choices have been made.
Generate an MCP from an APIA generator or framework
Interactively debug a connectionAn inspector or playground
Verify protocol conformanceA conformance suite
Review what agents receive and executeMCP Doctor
Trust follows the data
Know where every byte goes.
Inspect
Connects from your machine and writes locally. Nothing is sent to Coefficient.
Eval
Uses your credential. Task, schemas, calls, and results go to your selected model provider.
Telemetry
There is none in the CLI. The website has no product analytics or advertising pixels.
Dated public benchmark
State of MCP Quality 2026
- Observed
- 2026-07-10
- CLI
- 0.4.1
- Connected
- 5/10
- Rows scored
- 5
A reproducible methodology snapshot—not a claim that this catalog represents the entire MCP ecosystem.
Review results and failures →Current product / future product
The CLI is ready now. Hosted CI is not.
Use the free MIT-licensed CLI today. Join only if recurring private evaluation and regression checks would be useful to your team.
Read the CLI documentation →Questions before you run it
Plain answers.
- What does a passing eval prove?
- It proves that the selected model produced at least one real, non-error result from your MCP. It does not formally prove that every part of an arbitrary natural-language task was semantically correct.
- Do I need an API key?
- No for inspect and benchmark. Eval uses your own OpenRouter, OpenAI, Anthropic, AI Gateway, or Ollama credential. Credential values stay local; tasks, tool schemas, calls, and results go to the selected provider.
- Does it work with Stainless or Speakeasy output?
- Yes. MCP Doctor connects to the live MCP surface, whether it was generated, framework-based, or hand-written.
- Is it open source?
- Yes. The CLI is MIT-licensed. Hosted CI is not available yet.
- Who operates this?
- Roundtable Design Ltd (England and Wales, company no. 11924746), trading as Roundtable Studio. The optional waitlist stores your email and consent; product telemetry is off.
- Pricing?
- The CLI is free. There is no checkout. Paid hosted CI is not offered until that product exists.
Start with the live surface
See what your agents will see.
npx --yes @coefficient-work/mcp-doctor@0.4.7 listPublic GitHub and npm identity is coefficient-work. Install @coefficient-work/mcp-doctor; do not install the unrelated unscoped mcp-doctor package.