Claude

Claude plays two separate roles in Me:

1. The system’s model. It phrases messages to people, reads their free-text replies back into answers, classifies, and extracts fields from invoices. Provider com.me.model.anthropic.
2. Claude Code development sessions. Me runs Claude Code headless on a repository, in its own worktree and branch, and tells you when it needs you. Provider com.me.claude.

This page is about Claude inside Me. Claude Code using Me (the MCP server, the skill and the attention hook) is covered in Getting started.

Claude as the system’s model

me integration add com.me.model.anthropic --name Claude      # paste an API key or a subscription token at the hidden prompt
me integration add com.me.model.anthropic --name Claude --model claude-opus-5-5     # optional: another model

At the prompt, paste one of:

• An API key (sk-ant-api…), from console.anthropic.com (Settings → API keys). It is billed per use, through the Messages API.
• Your Claude subscription: run claude setup-token on any machine with Claude Code, and paste the sk-ant-oat… token it prints. Calls then run through claude -p on the server’s machine and count against your plan, so Claude Code must be installed there, for the user the server runs as (or set ME_CLAUDE_PATH on the plugin). At most two calls run at a time, each with a 60-second limit.

The default model is claude-sonnet-5-5. The check sends Claude one tiny request, so a bad key or token is caught at once.

What changes once it is connected:

• Messages to people are composed from intents. Inbox messages, reminders, nudges and the questions workflows ask are no longer fixed wording.
• Replies are understood. me conversation reply --text "Friday works best" fills in an ask’s answer, instead of coming back unclear.
• The system can answer texts and email from members (system’s number), and me money invoice ingest extracts invoice fields.

The bare model operations are also commands (me model generate, me model classify, me model extract), mostly for workflows.

Sensitivity. Everything sent to a model first passes the PHI egress filter. Without a signed BAA with Anthropic, nothing above Internal is sent to Claude, and requests about Confidential things (an unusual charge, for example) fall back to fixed wording. The operator raises the ceiling once a BAA exists, with Me:Coverage:com.me.model.anthropic (see the operator guide).

You see	What to do
“Anthropic refused the API key.”	me integration add com.me.model.anthropic with a working key
“No Claude integration is set up in this tenant.”	Add one (above)
“Claude could not be reached.”	Check the network; try again
“Claude Code (claude) is not installed on the server.”	Subscription token: install Claude Code for the user the server runs as, or set ME_CLAUDE_PATH
“Claude Code refused the subscription token.”	Run claude setup-token again and re-add the integration
“The Claude subscription’s usage limit is reached.”	It resets on your plan’s schedule; try later
“Claude Code did not answer within 60 seconds.”	Retry shortly

Claude Code development sessions

com.me.claude lets you, or an agent working for you, start Claude Code on a repository on the server’s machine:

me claude session start --connection conn_… --repo Me.CLI --task "Fix the flaky timer test" [--permission-mode plan|acceptEdits|default] [--max-turns 50]
me get cls_…                                            # state, last message, branch
me claude session respond --session cls_… --text "Yes, use the second approach"
me claude session stop --session cls_…
me claude session list

• Each session gets its own worktree at <workspace root>/.me-worktrees/<id>, on branch me/<id>.
• Starting and responding are High risk, so each waits for your approval, because a session runs code.
• You are told when a session needs you: when it asks a question, lacks a permission, fails or finishes. Events: claude.session.started, claude.session.needs_attention, claude.session.finished.

Setting it up (a person’s step):

me integration add com.me.claude --name "Dev" --workspace-root D:\src --credential-kind oauth_token
# credential: an API key (credential-kind api_key) or a `claude setup-token` token (oauth_token), at the hidden prompt

Claude Code must be installed on that machine (npm i -g @anthropic-ai/claude-code). Repositories must sit directly under the workspace root, and each needs at least one commit. The check runs claude --version and makes no model call.

Not in the local install yet. The plugin is built and on main, but deploy/local/install.ps1 doesn’t install it, so me integration list doesn’t offer com.me.claude on a local install. This has been reported.