One command, no project
Find out what a server’s tools actually do.
An MCP server can annotate a tool readOnlyHint: true. Nothing checks that. Agents use it to decide whether a tool can be called without asking you first.
RigorRun fetches the server, pins it to the exact bytes the registry published, runs it in a container with no network and no access to your machine, calls each tool with arguments derived from its own schema, and reads the filesystem before and after to see what actually changed. Then it compares that with what the server declared about itself.
rigorrun verify npm:@modelcontextprotocol/server-memory@2026.8.31 No project, no browser, no agent, nothing to configure first. Needs Docker, because every server runs in a container that is thrown away afterwards.
resolve npm:@modelcontextprotocol/server-memory@2026.8.31
stage installing with lifecycle scripts disabled
isolation measuring whether reset actually resets
discover 1 tool(s)
digest sha512-/EMH++1j2e+VmHwpibw1IUPHxA0tcQY09Yo5/OI…
isolation RESET
egress none
Declared, versus what it did
tool declared result verification
----------------- ------------------- ------- ------------
sequentialthinking readOnlyHint: true CONFORMS PARTIAL
sequentialthinking destructiveHint:false CONFORMS PARTIAL
sequentialthinking idempotentHint: true CONFORMS PARTIAL
Every discovered tool was exercised.What each exit code means
- 0 Verified
- Every claim it could check held, and nothing was contradicted.
- 1 Contradicted
- A declaration the server made about itself was contradicted by what it was observed to do.
- 3 Too little
- It ran, and established too little to be worth much. Fewer tools were reachable than you asked for.
- 2 Harness failed
- Something in the setup or the runtime broke. Never a finding about the server, and never dressed up as one.
Nothing it could not establish is reported as a fact. A tool it could not exercise safely is listed with the reason, and a record that omits what it could not reach is not a cleaner record — it is a false one. Run against 4 published servers, it exercised 19 of 37 tools; the rest are named rather than rounded away.