Working with AI

MCP Server

Exponential ships an MCP (Model Context Protocol) server so AI agents can manage issues with structured tool calls instead of shelling out to the CLI.

Setup

Project-local (recommended)

Create a .mcp.json at your repo root. This is checked into git, so every contributor gets the same tools:

{
  "mcpServers": {
    "xpo": {
      "command": "xpo",
      "args": ["mcp"]
    }
  }
}

User-level

Add it globally so it applies to every project:

claude mcp add xpo xpo mcp

Or manually add the snippet to ~/.claude/settings.json under mcpServers.

After any method, restart Claude Code (or run /mcp in-session).

Available tools

Tool	Description
`list`	List issues with filters (status, label, assignee, parent, text match)
`show`	Fetch one issue with description, dependencies, and comments. `include_spec`, `include_walkthrough`, and `include_events` inline more
`history`	Event audit trail for an issue
`add`	Create an issue with status, labels, parent, story points, links. `before`/`after` place it next to another issue
`update`	Patch any subset of fields. `links` replaces the full relationship set, `null` clears `assignee`, `cycle_id`, `priority`, `story_points` and `parent`, and `before`/`after` reorder
`comment`	Add a markdown comment
`link`	Add a relationship between two issues (bidirectional; one link per pair, `overwrite` replaces it; only `blocked_by` gates `start`)
`unlink`	Remove one relationship, whichever issue stores it
`start`	Start work: set to DOING, assign to the agent's principal, create a worktree. Returns the worktree path. Safe to call again from inside the worktree
`merge`	Merge a committed worktree into the default branch (squash by default), close the issue, and clean up the worktree
`spec`	Read, write, or delete an issue's design spec
`walkthrough`	Read, write, or delete an issue's implementation walkthrough
`artifact`	Attach, read, list, or delete other files on an issue
`rationale`	Search specs and walkthroughs for prior design decisions

Git tools

Read-only tools for what an issue changed and why code looks the way it does:

Tool	Description
`git_status`	An issue branch's state: ahead/behind the default branch, and staged, unstaged and untracked files in its worktree
`git_diff`	What an issue changed, from its merge base to its worktree, including uncommitted changes. `stat` for the file list only
`git_log`	Commits on an issue's branch, or on the default branch, optionally only those touching a path
`git_show`	One commit's metadata, message and patch
`code_history`	The commits and issues that touched a file or a line range. Pair it with `rationale` or `spec` to explain why code is the way it is

xpo builds every git command itself: the caller gives only an issue ID, a path, a SHA, a line range and a limit. External diff, textconv, fsmonitor and hooks never run, and the tools take no index locks. Output is capped at 64 KiB and marked truncated: true.

Reduce permission prompts

Allow-list the xpo tools in .claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__xpo__*"
    ]
  }
}

All xpo changes are git-tracked, so any mistake is recoverable with git.

PreviousAgent Configuration NextScripting with JSON