Give each agent a role, choose how they take turns, and let the real CLIs on your machine work on your code, or any question, and answer each other.
$ git clone https://github.com/Agentic-System-Lab/harness-relay $ cd harness-relay && npm install && npm run dev
A new group, then Auto team on this repository, from the request to the answers and a search. Sped up while the agents work.
Harness Relay never calls a model API. It starts the claude, codex and
grok you already have, with their own tools, sessions and logins, passes each agent the
messages it hasn’t seen yet, and streams everything into one chat.
claude with stream-json in both directions; its permission requests become approval cards.codex app-server: threads, turns and approval requests.grok agent stdio.You don’t need all three. Any subset works, and you sign in to each one from Settings.
Each conversation has its own rules: who speaks, in what order, and when it ends. Address anyone with an
@mention and only they reply.
Agents speak in order, and each sees what earlier speakers said. Set one to five rounds; the run ends
early when everyone in a round replies [PASS].
Good for discussions, reviews and refining an answer.
Everyone answers at the same time, on their own. Later rounds see the earlier answers. With edit access each agent works in its own git worktree, and you apply the result you like.
Good for comparing approaches and getting independent opinions.
One agent at a time. An agent hands the floor to a teammate by @mentioning them; if it
mentions nobody, the floor returns to you. You set the maximum number of hand-offs.
Good for pipelines: plan, implement, review.
Design it yourself. Give each agent a position (architect, builder, reviewer, skeptic), pick its model and reasoning effort, and set the speaking order. Up to eight members, several on the same CLI if you like. Roles can change mid-chat.
Or use Auto team. Describe what you need. The lead agent reads it with every template in full, then uses one as it is, adapts one, or designs a new team, and says why. You confirm before anyone answers, and can keep its team as a template.
The lead agent picks or designs the team for what you describe.
Claude plans, Codex builds, Grok reviews, handing off with @mentions.
Propose, challenge, converge. Good for technology choices and design discussions.
Three agents solve the same problem independently at the same time, for easy comparison.
Not just code: analyse, fact-check and combine into one clear answer.
The Navigator directs, the Driver writes the code, QA writes tests and verifies.
Agents work with the permission you choose for the conversation. Anything beyond it, they ask you in the chat, and they wait.
npx tsx -e "import { mentionsInMarkdown } from './src/shared/mentions.ts'; …"
Allow once, always allow it in this chat, or decline with a note on what to do instead.
Before I compare the options: does the old /v1 API have to keep working, or may it break?
When a good answer depends on something only you know, an agent asks and the team waits. Once you reply, that agent answers first, with your answer in hand.
The working directory becomes a git repository, and each reply lists the files that agent changed. In parallel mode every agent has its own worktree until you apply one result.
A message sent mid-reply waits above the box and reaches the next agent before its turn. Edit or withdraw it until then, or stop the team and send it now.
Interface excerpts recreated from the app, with sample content.
Nothing is hosted. The server listens on your computer only, and agent replies cannot make your browser load remote content.
127.0.0.1; cross-site requests are rejected.| Read-only | Reading, searching and looking things up on the web. |
|---|---|
| Edit | Also editing files in the working directory and running project commands such as tests and builds. Pushing and publishing are not included. |
| Full access | Everything, without asking. Use it only in directories you trust. |
Edit doesn’t sandbox project code: tests and scripts run code the agents can change. For code or instructions you don’t trust, use read-only.
claude),
Codex CLI (codex), Grok CLI (grok)
$ git clone https://github.com/Agentic-System-Lab/harness-relay $ cd harness-relay $ npm install $ npm run dev
Open http://localhost:5173. Under Settings & sign-in, check that each agent is signed in. Then New chat: choose Auto team, design your own, or pick a template.
One process instead: npm run build && npm start serves everything on http://127.0.0.1:8787.