Privacy and security for external agents

Business Summary: External agents run on your account and your computer, so their keys and files stay with you. Harmony moves your requests and shows the results; it never stores your agent's password or API key.

What Harmony sees and stores

Harmony stores your agent setups (the agent, its name, its computer and folder), the requests you send, the steps each agent reports (commands it ran with their output, and which files it changed), and the agent's answers. These appear in your agents panel and stay private to you unless you share an answer with Collaborate.

Harmony does not store your agent's API key, your Claude or ChatGPT sign-in, or the files on your computer. Those stay on the computer that runs the agent. When you choose an option under Authentication, Harmony only records which kind of sign-in you use.

Keys you add yourself in the Model provider keys card in Profile → External agents are different. Harmony holds those so its gateway can call the provider for you: they are sealed when saved, never shown again, and never placed in a prompt, a computer or the code sandbox. An Anthropic, OpenAI, Google Gemini or OpenRouter key can also answer your own chats and Loops when you leave Use for chat and Loops on. See Bring your own model.

How this fits in Harmony

flowchart LR
  subgraph Yours[Your computer]
    Runner[Harmony runner]
    Agent[Agent and its sign-in]
    Files[Your files]
  end
  Harmony[Harmony]
  Runner -->|asks for work| Harmony
  Harmony -->|your request| Runner
  Runner --> Agent
  Agent --> Files
  Runner -->|steps and answer| Harmony

The runner only reaches out from your computer; nothing connects in. Each computer has its own code, which Harmony keeps only in a scrambled form that cannot be turned back into the code. Removing the computer, or choosing Pair again, makes the old code useless straight away.

Who can do what

Only you can ask your agents, because they act as you. Teammates cannot send requests to your agents or see your agents panel. They see what you share with Collaborate, and pull request review cards. Reviews they submit use their own GitHub accounts.

What agents are allowed to do

An agent can do on its computer what you could do there: read and edit files in its working folder, run commands, and use the internet. Give it a working directory for the project it should touch, and keep secrets you would not want it to read out of that folder. Agents running in headless mode accept their own file edits without asking, so review their pull requests before merging.

On a Harmony cloud computer, agents run inside an extra security sandbox, separate from every other person's computer and from Harmony systems.

Ready to try it?

See Connecting a computer to set up the runner, then Adding an external agent.

Example scenario

An administrator asks whether Claude Code in Harmony could leak the company's Anthropic key. The answer: Harmony never receives it. The key stays in Claude Code's settings on the engineer's laptop, and Harmony only ever sees the requests, steps and answers, all private to that engineer until shared.

FAQs {#faqs}

Q: Can Harmony staff read my agent conversations? A: They are stored in your organization's Harmony data like other private chat, under the same access controls and policies.

Q: What if my install command leaks? A: Choose Pair again on that computer. The old code stops working immediately and you get a new command.

Q: Can a teammate run my agent on my plan? A: No. Only you can ask your agents.

Q: Does the agent see my whole computer? A: It starts in the working folder you chose, but it runs with your user's permissions, like any program you start. Point it at the project folder and treat its changes like a colleague's.

Q: Is command output stored? A: Yes, the tail of each command's output is stored with the step so you can see what happened. Avoid having agents print secrets.