RFC: admin tools in the MCP server (organization and server level)

Summary

GlitchTip's MCP server (/mcp) lets an AI agent read issues, events, performance data and logs, and resolve or ignore issues. It has no administration tools.

This is a request for comment: what admin tasks would you want an agent to do through MCP? Reply with the concrete questions or tasks you would hand an agent.

Today, agents doing admin work fall back on worse options:

  • a full-scope API token pasted into a chat, with the agent calling the REST API directly;
  • a shell inside the container (manage.py shell);
  • clicking through the Django admin.

A small set of typed MCP tools would be narrower, use the normal OAuth login (including MFA), and could log every call.

Current lean

Minimal and read-only at first, at two levels:

Organization level, on the existing /mcp, for owners, managers and admins of an organization. These users can already see all of this in the UI or API. Possible tools:

  • members, roles, teams and pending invites;
  • projects with their DSNs;
  • alert rules and uptime monitors;
  • "can user X see project Y, and why or why not";
  • quota and usage.

Server level, for instance operators (superusers), on a separate endpoint that is off by default. These are support lookups that span organizations:

  • find a user by email and list the organizations they belong to;
  • show an organization's owner, members, plan and whether it is accepting events.

Planned constraints:

  • Account metadata only. Server-level tools never return issue, event or log contents from other organizations.
  • No generic tools. No "query any model" tool, no SQL, no management-command runner. Each tool answers one typed question.
  • OAuth only for server-level tools. Long-lived API tokens are refused there, and superuser status is checked on every call.

Not in this proposal

Write tools, such as:

  • creating a team or project;
  • inviting a member;
  • adding an alert.

These are the obvious next step for "have an agent set up GlitchTip for my organization", and we'd like to hear if you want them. The MCP server also returns event text that anyone with a DSN can submit, so an agent with write access can be steered by text injected into an event. Writes that grant access or send data out (member invites, role changes, webhook alerts, SSO settings, API tokens) would need an explicit human confirmation in the MCP client.

Timing

No implementation is scheduled yet. We are evaluating django-bolt for the web server, and its MCP integration can turn existing API endpoints into MCP tools that share the same permission checks. That would change how these tools are built, so this will likely wait until that decision is made. Feedback now shapes what gets built either way.

Questions

  1. Which admin tasks would you use? Concrete examples help most.
  2. Self-hosters: would you enable server-level tools on your instance?
  3. Do you need write tools, and which ones first?
  4. Any concerns about an agent having admin read access at all?

Related: #432 (Add Social App via API).

AI disclosure: drafted with Claude Code (Opus 5.5), which also searched the existing issues and reviewed the current MCP and permission code.