For agents

mxr already works with agents today through the CLI and the skill file.

That is the current story. It is simple, it is scriptable, and it does not pretend an MCP server exists when it does not.

Current agent surface

Surface	Status	Notes
CLI	shipped	structured output, batch ops, dry runs
Skill	shipped	documents the CLI for coding agents
Daemon socket	shipped	available for custom clients
First-party MCP server	not shipped	still on the roadmap

Why the CLI works well

• --format json gives structured output
• --dry-run previews risky mutations
• --search lets one command target a set of messages
• mxr history shows persisted mutation history

This is enough for a lot of useful agent work without inventing a new tool surface first.

Good patterns

Search first

mxr search "is:unread" --format json

Preview before changing anything

mxr archive --search "label:notifications older:30d" --dry-run

Check what happened after

mxr history --category mutation

Example workflows

Inbox triage

1. Search unread mail.
2. Read selected messages with mxr cat.
3. Draft replies or export threads.
4. Use --dry-run for any batch mutation.

Meeting prep

1. Search by sender and date range.
2. Export the relevant thread as markdown.
3. Summarize open items outside mxr.

CI cleanup

1. Search build notifications.
2. Group by thread or sender.
3. Preview archive/trash actions with --dry-run.
4. Apply mutations and verify with mxr history.

Safe defaults

• Prefer --dry-run on batch changes.
• Use --yes only when the workflow is already known and reviewed.
• Treat trash, spam, and unsubscribe as high-friction commands.
• Keep in mind that agent-safe permission presets are not shipped yet.

Current limits

• No first-party MCP server yet
• No read-only or draft-only agent mode yet
• No agent-specific account scoping yet
• No explicit send-approval flow yet

If you need those controls right now, treat mxr as a CLI tool that an agent can use carefully, not as a permissioned agent platform.