External agents

Business Summary: External agents let you bring the coding agents you already pay for, such as Claude Code, Codex and Cursor, into your Harmony workstreams and rooms. You hand them work like a teammate, watch every step they take, and share the results with your team.

Note: External agents are rolling out. Pairing your own computer, adding agents, asking them in chat, watching their steps, sharing replies and reviewing their pull requests all work today. Harmony cloud computers are available where your administrator has turned them on, and agents on a cloud computer need you to sign in to their service there before they can use your plan.

Many teams already have subscriptions to AI coding agents. Those agents are powerful, but they live in a terminal on one person's laptop, so the rest of the team never sees what they did or why. External agents fix that: the agent keeps running on your account, and the work shows up where your team already talks.

An external agent is different from Harmony Agent, the agent built into every workstream. Harmony Agent is part of Harmony and works with your connectors and company knowledge. An external agent is a program you run yourself, on your own account or plan, on a computer you choose. Harmony never sees that program's keys or password. It passes your requests to the agent and shows you what the agent did.

How this fits in Harmony

You set up external agents once, in your profile. After that they appear in every chat you use, ready to be asked.

flowchart LR
  You[You in a workstream or room]
  Panel[Agents panel]
  Computer[Your computer or a Harmony cloud computer]
  Agent[Claude Code, Codex, Cursor and more]
  Room[The room sees shared results]

  You -->|"@Claude Code ..."| Panel
  Panel --> Computer
  Computer --> Agent
  Agent -->|steps and answer| Panel
  Panel -->|Collaborate| Room

There are three parts. A computer is where agents run: your own Mac or Linux machine, or a Harmony cloud computer. An external agent is one coding agent on that computer, with a name and a purpose. The agents panel is the side panel in every chat where you talk to agents and watch their work.

What you can do

You can ask an agent from any workstream by typing @ and picking it, for example @Claude Code fix the failing login test. The agents panel opens on that agent and the work happens there, privately. You see each command the agent runs as a terminal block with its output, and each file it changes. When it finishes, its answer appears in the panel.

If the answer is useful to the team, choose Collaborate to share it with the room as a post everyone can edit. If the agent opens a pull request, the chat shows a review card so teammates can approve it or request changes under their own GitHub accounts. If the agent uses a web browser, the Browser button in the agents panel shows you what it sees.

Articles in this section

• Connecting a computer: pair your Mac or Linux machine so agents can run on it.
• Adding an external agent: choose Claude Code, Codex, Cursor or another agent and set it up.
• Working with agents in chat: ask, watch, stop and share.
• Harmony cloud computers: computers Harmony runs for you, for when your laptop is closed.
• Reviewing pull requests from agents: approve or request changes from the chat.
• Privacy and security for external agents: what Harmony sees and what it never stores.

Example scenario

Priya leads a small product team. She pairs her laptop, adds Claude Code, and opens the team's Checkout workstream. She types @Claude Code the login test fails about one run in ten, can you find out why?. In the agents panel she watches Claude Code run the test twenty times, find a race, change one file, and open a pull request. She chooses Collaborate so the team sees the explanation, and her colleague Sam approves the pull request from the review card in the chat.

FAQs {#faqs}

Q: Do I need a paid plan with the agent's company? A: On your own computer, yes. The agent runs on your own account with that company, such as a Claude or ChatGPT plan, or an API key you set up on that computer. On a Harmony cloud computer, the agent reaches its model through Harmony instead: your own provider key comes first, then a key your organization added, and without either the agent uses Harmony's models. See Bring your own model.

Q: Which agents can I use? A: Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, Goose, Amp, Cline and many more. You can also add any agent you start from a command, an MCP server, or an API endpoint. See Adding an external agent.

Q: Can my teammates use my agent? A: Not yet. An agent acts as you, on your plan, so only you can ask it. Teammates see what it shares to the room and can review its pull requests.

Q: Is this the same as attaching Harmony Agent? A: No. Harmony Agent is built in and works with your connectors and company knowledge. External agents are coding agents you run yourself. Both live in the same agents panel, and you switch between them at the top of it.

Q: Does my computer need to stay on? A: For agents on your own computer, yes: it needs to be awake and online when you ask. A Harmony cloud computer starts on its own when an agent is asked.