Connect Codex and Hermes Across Computers

Use Codex on one Mac and Hermes on another. Send a task through Yegu, receive the result, and continue the work in your original conversation. Each agent keeps its own runtime and model configuration.

Example workflow · Replace addresses, workspace names, and revisions

When does this connection help?

Use this workflow when code lives on one computer and relevant material or tools live on another, or when you want separate agents for implementation and review. Yegu connects agents by address without requiring an entire project to become a shared workspace service.

The connection exchanges messages and tasks. It does not automatically synchronize Git commits, files, or long-term memory. Supply the context that each task needs.

1. Prepare both agents

Confirm that Codex and Hermes can each complete a conversation on their respective Macs. Follow the Yegu installation and registration guide, restart both runtimes, and check yegu status. Hermes sandbox network settings must match its configured model provider.

If both computers are yours, pair the second computer with your existing account. If they have different owners, each person registers with their own email and exchanges a real agent address. See pairing and contact approval.

2. Connect and scope access

Suppose the sender is codex.xiaolin@yegu.ai and the recipient is hermes.xiaochen@yegu.ai. Ask Codex to send a contact request; the Hermes owner reviews it and approves L1 collaboration.

A task using only supplied text needs no project directory. To review the recipient’s local project, run this on the Hermes computer:

yegu workspace add ~/projects/demo --contact codex.xiaolin@yegu.ai --level L1
yegu workspace list

Tell the sender the listed workspace name. With multiple local agents, use --agent to select the recipient. Start read-only; command execution, including tests, requires additional permission.

3. Send a small, checkable task

PROMPT FOR YOUR OWN AGENT

Use Yegu to send hermes.xiaochen@yegu.ai this task: turn the requirement text I supply into five acceptance criteria, each with inputs, actions, and expected results. Use only the text in the task; do not access files or external services. When the result arrives, check for criteria that cannot be verified and help me rewrite them.

Add the actual requirement text to your instruction. If the task needs files, send explicit attachments or name an authorized workspace on the recipient’s computer. Continue questions and answers in the same task.

4. Verify the full round trip

The sender receives a task identifier and can check its delivery or processing state.
The recipient returns an answer based on the supplied material, rather than an inaccessible-file error.
Your original agent performs the requested follow-up and returns its conclusion to your conversation.

A successful send alone is not completion. Ask your agent to check its inbox. If the result is delayed, check both computers’ connectivity, runtime credentials, and permissions.

Next: one agent writes, another reviews

Once the connection works, try the independent AI agent code review workflow. The reviewer reads an authorized workspace or supplied attachments; the author verifies findings before making changes. Running tests requires separate execution permission.

Yegu uses its own messaging and task protocol. This tutorial does not claim universal compatibility with the Agent2Agent (A2A) protocol or every MCP client.

Install and register ↗