Skip to main content
The Mzizi MCP server gives agents the registry, the brand tokens, the architecture model, the doctrine, the agent skills and these docs over the Model Context Protocol. It is the mzizi-mcp Cloudflare Worker, published as @nyuchi/mzizi-mcp (0.9.1, read from npm on 29 September 2026).

Connecting

Streamable HTTP. mzizi.dev/mcp answers 308 to this endpoint, so old clients keep working, but point new ones here directly. To run it locally over stdio, with no network for anything but the docs tools:
The stdio build takes no environment and holds no credential. The registry data and skills ship inside the package.

Who can use it

Rolling out. The owner decided on 29 September 2026 that the MCP server is free with no gate. Every tool except the Fundi tools is to be served without sign-in. On 29 September 2026 the hosted endpoint still answers an unsigned request with 401 {"error":"invalid_token"}, because the change is landing in the tooling repository now. Until it deploys, use the stdio build above, which has never needed sign-in.
The intended behaviour: The Fundi tools stay gated because they tie into the console at app.mzizi.dev: mzizi_report_issue files into the Fundi issue desk, and mzizi_fundi delegates work to Fundi on a user’s behalf. Both need to know who the user is. The server advertises its OAuth metadata at /.well-known/oauth-protected-resource, so MCP clients can run the sign-in flow for those tools themselves.

Discovering the tools

Ask the server, not a page. A list written down anywhere, this one included, is only as fresh as its last edit.
  • tools/list over MCP is the live answer.
  • https://mcp.mzizi.dev/catalogue.json is the same list over plain HTTP, with no sign-in, including which older tools each one replaces.
  • The mzizi_mcp_describe tool, and the mzizi://tools resource, describe the tools and the data sources.
On 29 September 2026, catalogue.json listed fourteen tools: Start with mzizi_search when you do not know the name of the thing you want, and mzizi_get_component when you do.

Where its data comes from

Nothing is read from a database, and no registry data is fetched at request time. Because the generator runs the registry’s route handlers rather than re-deriving their output, a tool answers with the payload api.mzizi.dev/v1 would give at the same commit. catalogue.json names the pinned registry commit in source.registry. Moving to newer registry content is a change to that pin, so it is a reviewed commit. The server holds no database credential. Former nyuchi-* component names resolve to their mzizi-* names through the registry’s rename map, bundled with the rest.

The docs tools

These docs run their own MCP server, which Mintlify hosts at https://docs.mzizi.dev/mcp. It is public and needs no sign-in. The Mzizi MCP server lists its tools under a docs_ prefix and forwards each call unchanged:
  • it caches the upstream tool list for five minutes;
  • if a refresh fails it reuses the last good list, and with nothing cached it falls back to a built-in snapshot of the upstream definitions;
  • a failed call comes back as a tool error naming the docs server and the reason.
The docs_* names mirror whatever docs.mzizi.dev/mcp lists, so they can change when these docs change. You can also connect to https://docs.mzizi.dev/mcp directly.

The Fundi tools

Fundi is the self-healing agent behind the console, run under Nyuchi. Two tools reach it. mzizi_report_issue reproduces and drafts an issue, then logs it to Fundi’s issue desk. Fundi files the GitHub issue with its healing plan, and records the issue’s lifecycle, which is how the desk and the console can tell you whether your report was picked up. mzizi_fundi delegates long runs, such as a security, chaos or accessibility run, to Fundi over the Agent2Agent (A2A) protocol. A run is a task with a lifecycle, not a blocking tool call: you submit it and get a task id back at once, then poll its status or cancel it. It replaces the older fundi_status, fundi_submit_test, fundi_task_status and fundi_cancel_task tools.
The A2A bridge is built through its second stage: the agent card, task submission, status and cancellation. The runs behind it, streaming updates and push notifications are still design. A submitted task can come back parked as accepted but not executed.
Both tools act for a real, signed-in console user. The server never hands a machine credential to a client or to the CLI.

Use cases

Building against the registry

  1. mzizi_list_components, filtered to the node you are building at.
  2. mzizi_get_component for the full document of anything that looks right. Where a Rust implementation exists, prefer it (see Mzizi Roots).
  3. Install it (see consuming the registry).

Reviewing a change

  1. mzizi_get_tokens to check a component uses published tokens, not raw values.
  2. mzizi_get_architecture to check it sits where its imports say it does.
  3. mzizi_check_accessibility for contrast on any new colour pairing.

Answering a question about Mzizi

docs_search_mzizi, then docs_query_docs_filesystem_mzizi to read the page it found.