MCP Playground: Run Your Agents' Tools Live, Right From the Docs
Last week we made MCP tools a first-class API contract: point Archyl at a live MCP server, and it discovers every tool and its input schema, documented and linked to your C4 model like any REST or GraphQL contract.
That answered the first question: what can my agents do?
Today we're answering the second one: does it actually do that?
Meet the MCP Playground — a new tab on every live MCP contract that lets you invoke any tool against the real server and inspect the real result. No client setup, no JSON-RPC by hand, no leaving your architecture docs.
From schema to form, automatically
Open a live MCP contract and you'll find a Playground tab next to the schema. Pick a tool from the list — the same searchable list you documented — and Archyl turns its inputSchema into a form:
- Strings, numbers and booleans become typed fields.
- Enums become dropdowns with the allowed values.
- Required parameters are marked, and validated before anything is sent.
- Nested objects and arrays get a raw JSON field, validated as you submit.
Fill it in, hit Run tool, and the result comes back with its duration — rendered the way results deserve to be rendered. A Formatted view shows the response as clean, pretty-printed JSON. A Tree view makes large payloads navigable: collapse what you don't care about, expand what you do. Text, images, and resources returned by a tool all render natively.
If the tool returns an error, you see exactly what an agent would see — the error flag and payload, surfaced in red, not swallowed.
Your token never leaves your browser
The Playground follows the same security model as live discovery, and it's worth repeating because it's the whole point:
Every call runs from your browser. When you hit Run tool, your browser talks to your MCP server directly. Archyl's backend is not in the path.
- The token is never stored. You enter it for the session; it lives in the page and nowhere else.
- Results are never persisted. What comes back is rendered, read, and gone when you close the tab.
- No server-side reach into your network. Because the request originates from your machine, the playground reaches
localhostand private servers — and there is no server-side request path to abuse.
The one trade-off, as with discovery, is CORS: the target server must allow Archyl's origin. For servers you control, that's one line of config.
Why this matters
MCP tools are the interface your AI agents actually use. Until now, checking one meant wiring up a client, crafting a JSON-RPC envelope, or just… trusting the description.
Now the documentation is the test bench:
- Reviewing an agent surface? Run the tools and see real shapes, not described ones.
- Debugging why an agent failed a call? Reproduce it in two clicks with the exact same arguments.
- Onboarding someone to a service? Send them the contract — they can read every tool and try it.
This is what we mean by living documentation. A spec that can be executed doesn't drift quietly; it gets caught the moment it lies.
Dogfooded, as always
Archyl is itself an MCP server — 181 tools as of this week. The Playground's first user was our own contract: we run list_projects, get_project_c4_model and friends against our own endpoint, from our own docs, every day. The screenshot above is exactly that.
Try it
Open a project → API Contracts → any MCP contract with a live endpoint → Playground. Enter your token, pick a tool, run it.
Your architecture docs just learned to execute.