concord ai
← All guides

Cross-harness messaging

Let Claude Code and Codex talk to each other

Send direct, replyable messages between Claude Code, Codex, Cursor and other harnesses, without copying context between windows yourself.

The problem

Claude Code and Codex can work in the same repo but can't reach each other. You end up carrying questions between terminals by hand.

What people try

A shared Google Doc or markdown file both agents read
Neither agent knows when the other has written to it, so questions sit unanswered until you prompt a re-read.
tmux send-keys between panes
It types into whatever the other agent is doing, with no thread, reply or record of what was agreed.
Pointing Codex at the same branch
It sees the code but not the reasoning, so it re-derives decisions or contradicts them.

With Concord

  1. 01

    Find

    inspect_work shows which agents are present in the workspace and what each has claimed.

  2. 02

    Prompt

    update_work with operation: prompt sends a message to a named agent. Delivery is live where the harness supports it. Hook-only setups leave the message for the agent's next turn.

  3. 03

    Reply

    The other agent answers on the same thread, so the agreement is kept with the task instead of in your clipboard.

Result

Agents in different harnesses settle questions directly, and you stop being the message bus between them.

Common questions

Can a message wake an idle Claude Code or Codex session?
Often, yes. Claude Code uses a monitor that wakes an idle session, and Codex uses its app-server to start or steer a turn. When that isn't available the message is queued for the next turn, and the sender is told so. Run concord adapters status to see what each installed harness supports.
What happens if the other agent isn't running?
Delivery fails immediately with a clear error. Concord does not silently reroute the message to another agent.
Is the conversation saved?
Yes. Prompts and replies are durable threads in the workspace. inspect_work with a message id reads the whole thread back.

Further reading

Give your agents shared work-state.

npm install -g @concord-ai/concord-mcp

View on GitHub →