claude-message

claude-message wraps claude -p --output-format json and uses agent-message as its transport layer.

Install:

npm install -g agent-message claude-message

claude-message expects both agent-message and the claude CLI to already be available on your PATH.

If --to is omitted, claude-message uses the current agent-message master value.

Behavior:

1. Starts a fresh agent-{chatId} account with a generated password.
2. Sends the --to user a startup message with the generated credentials.
3. Reuses the Claude session_id for the DM session and resumes later turns.
4. Watches agent-message DMs for plain-text requests, adds a 👀 reaction to each accepted inbound DM, and passes the request to Claude with explicit instructions to send the final user-facing result directly with agent-message send --from agent-{chatId}.
5. If Claude fails, the wrapper sends a failure json_render notice itself.
6. After a successful turn completion, replaces the inbound 👀 reaction with ✅.

claude-message now follows the same delivery model as codex-message for successful turns: the agent is expected to send the final result itself. The wrapper still handles startup, reactions, and failure notices.

Example:

agent-message config set master jay
claude-message --model sonnet --permission-mode accept-edits
claude-message --to alice --model sonnet --permission-mode accept-edits
claude-message upgrade

Build from the repo root:

make claude-message-build
./claude-message/target/debug/claude-message --to jay --model sonnet

Useful flags:

• --to <username> overrides agent-message master
• --cwd /path/to/worktree
• --model sonnet
• --permission-mode accept-edits
• --allowed-tools Read,Edit
• --bare

Notes:

• claude-message always passes --dangerously-skip-permissions to Claude.
• The wrapper therefore does not pause for Claude permission approvals.