Guide · Deployment and agents

MCP ticketing: let an agent draft tickets safely

MCP ticketing gives an AI agent tools to draft, review and route tickets through Relay. The safety rule is simple: reads work with a scoped key, but every external write needs confirm: true on the call. An agent can prepare work, yet it cannot push a ticket to a tracker without that explicit parameter.

By the Relay maintainers · Last updated 19 November 2026 · Source on GitHub

What is MCP ticketing?

MCP is a protocol that lets a model call external tools. Relay runs an MCP server, in both stdio and streamable HTTP modes, as a thin client of its REST API. That means an agent can ingest a note, list drafts, explain an assignment or search memory using named tools instead of hand-written API calls. The server exposes 28 tools, and the same review and audit rules apply whether a person or an agent calls them.

How does the confirm gate work?

A tool call that pushes to a tracker, changes a policy or connects an integration is an external write. Relay requires confirm: true in that call before it proceeds. Without the parameter, the write is refused. This is enforced in the service layer, not only in the tool definition, so a different client cannot bypass it. A reviewer can give an agent a viewer key to remove write access entirely.

What can an agent do with the tools?

The tools are grouped by job. Capture tools ingest and edit drafts; review tools approve, reject, delegate and push; assignment tools suggest and explain owners; memory tools search decisions and commitments; policy tools read and propose changes. Each group maps to the same services a human interface uses, so there is no separate, weaker path for agents.

MCP tool groups and their jobs
GroupExample actions
CaptureIngest a note, edit or merge a draft
ReviewApprove, reject, delegate, push
AssignmentSuggest and explain an owner
MemorySearch decisions and commitments
PolicyRead policy, propose a change, dry run

How does an agent connect?

  1. Create a member and key. Issue a dedicated scoped key for the agent.
  2. Generate the config. relay agent-config prints the MCP client settings.
  3. Store it safely. Keep the config with restricted file permissions; secrets are masked in output.
  4. Start with reads. Ask the agent to draft tickets from a transcript and show what needs review.
  5. Confirm writes explicitly. Only pass confirm: true for a push you have checked.

What are the limits?

Agent actions are only as good as the source material and the policy around them. Extraction on messy multilingual fixtures measures about 0.78 precision and 0.75 recall, so drafts need review. Live model output varies between runs. Hosted device login and OAuth apps for agents are designed but not built; the CLI plus scoped keys is the working path today. Hosted plans are planned, not live.

Next step

Read the self-hosting guide to stand up the API and MCP server, then try a read-only agent against a scratch workspace.


Self-host Relay Read the auth guide