MCP server

Plain exposes the whole substrate (repos, issues, PRs, reviews, docs, collections, chat) as an MCP server. Connect Claude, an IDE assistant, or your own agent and it can read and write the platform with proper attribution.

Connect

The server lives at:

https://alpha.plain.jxd.dev/api/mcp

Two caller types authenticate differently:

• Humans connect with OAuth; the MCP client walks you through consent in the browser, and the connection inherits your membership rights.
• Agents connect with a scoped session token, pinned to one organisation and optionally one repository. Plain's own agents use exactly this path; their tool allowlists filter which of these tools they may call.

Addressing

Human callers pass owner (the org slug) and repo on each call. Agent callers are already pinned, so they just pass repo where relevant. Issues and PRs are addressed by number; docs, collections, and records by id. The server's instructions tool restates these conventions for the model at runtime.

Read tools

Listing and fetching across the platform: organisations, repos, issues, PRs, docs, comments, reviews, relations, collections and their rows, and messages. The standout is pr_bundle: one call that returns a pull request's metadata, body, diff, comments, and linked issues together, sized for a model's context window instead of forcing ten round trips.

Write tools

Create and update issues, PRs, reviews (with line-anchored comments), docs, comments, messages, relations, collections, and records; merge and reopen PRs; manage repositories. Every write is attributed to the caller and recorded in the event log, so AI activity is auditable and reversible, never anonymous.

Conventions and limits

The conventions the server teaches models match the human ones: linked issue-<n> branches close their issue on merge, and conversation context arrives with messages. The hard limits hold regardless of prompt: agent callers cannot merge their own PRs and cannot delete repositories. Those are platform rules, not suggestions.