Guide: Connector automation

muaz speaks the Model Context Protocol (MCP), so any MCP server becomes a set of tools an agent can call. That turns muaz into an automation surface over the systems you already use — with the approval policy as the guardrail.

Add a connector

Open the UI Connectors page and pick one from the catalog (or paste an { "mcpServers": { … } } config). Core-tier connectors — GitHub, filesystem, fetch, memory, Linear — are the small set muaz CI-tests against real servers; see the Connectors guide. Fill in any required ${VAR} values (a token, a path), and use Reconnect to verify it comes up.

MCP runtimes (Node/uv) are downloaded on demand the first time a server needs them, so the first connect may take a moment.

Attach it to an agent

A server is opt-in per agent: list it under the agent’s tool access and give its tools an approval level (off / ask / auto). A good default is ask for anything that writes and auto for reads.

Automate

Now ask the agent to do the work:

• “Open a GitHub issue summarising the failing tests, and link the PR.”
• “Find every Linear ticket assigned to me that’s been stale for two weeks and draft a status update.”
• “Read ./incidents/ and produce a weekly summary.”

Each tool call that isn’t auto prompts you first — you approve the exact action (and can Approve for session to stop being asked for a repeated call).

Keep it safe

• Least privilege: attach only the servers an agent needs, and keep writes at ask.
• Egress: an agent with network: none refuses HTTP MCP entirely (stdio MCP children are outside that boundary — install only trusted, signed connectors).
• Audit: every resolved tool call is recorded in the audit log (muaz audit tail) with the agent, tool, an argument hash, and the decision.

Full MCP reference and the pasted-config format: the Connectors guide and Tools & approval.