The MCP Playground
You can read a tool list and still have no idea how an agent will behave against it. Descriptions that look fine to a human turn out ambiguous to a model, a base URL points at the wrong environment, a required parameter is never filled in because nothing said it mattered. The MCP Playground is where you find that out, before you share the install link.
The Playground connects a model to your MCP tools so you can inspect its arguments and results. Calls can reach your real API and have real side effects.
Starting a session
Open the MCP server's Playground and choose an available model. Start with a read-only prompt whose expected result you can verify, such as asking for one test order by ID.
Supply the upstream credential when prompted. For OAuth authorization code or OpenID Connect, complete the browser sign-in and return to the Playground. Use Reconnect after an expired or changed authorization, or Clear to remove the current credential before testing another connection.
Reading a tool call
Every tool the model invokes is listed in the conversation, and each one opens up. That inspection view is the point of the Playground: it shows the arguments the model actually constructed next to the response your API actually returned. Most first-run problems announce themselves the moment you look at that pair.
The model picked the wrong tool
The model picked the wrong tool
Two tool descriptions overlap, or one is vague enough to look like a match for anything. Agents choose by description, so the fix is in the wording of the operation, not in the schema.
A required argument is missing or invented
A required argument is missing or invented
The parameter has no description, or its schema never says what a valid value looks like. A model given no example will produce a plausible one, and plausible is not the same as correct.
Every call returns 401 or 403
Every call returns 401 or 403
The Authorization tab is misconfigured, or the credential you supplied lacks the scope the operation needs. Clear the credential and re-enter it to rule out a stale value.
Calls that should work return 404
Calls that should work return 404
The server is pointed at the wrong base URL. This is most often production settings aimed at a staging host, or the reverse, and a base URL override fixes it without touching the spec.
The call succeeds but the model misreads the result
The call succeeds but the model misreads the result
The response schema does not describe what actually comes back. The agent is reasoning from the promise rather than the payload, so correcting the schema is what fixes the behavior.
The Playground calls your real API. It is a sandbox for the agent, not for your backend, so a tool that writes data will write data. Point the server at a staging base URL, or stay on read-only tools, before you let a model explore freely.
Access and lifetime
Editors and admins can use the Playground. Sessions last up to 24 hours, and a server must be published before it is callable. After deployment settings change, earlier sessions can become read-only. Start a new session against the current deployment.
Where to go next
AI readiness
Turn the problems you just watched happen into a scored list of schema and description fixes.
The public install page
Once the Playground is clean, this is the link you hand to everyone else.
If authentication fails, reconnect and verify the upstream account's permissions. If the tool list is stale, check the published deployment and start a new session. Clearing a Playground credential does not revoke that credential at the upstream provider.
Compare models — Preview
Choose between two and six available models to compare the same prompt with the same MCP connection and tool surface. Inspect each model's tool choices, arguments, results, and errors. Models execute separately, so a write tool can run more than once. Use read-only operations or isolated test data.
On this page
- The MCP Playground