Harnesses
Use Claude Code, Codex, Pi Agent, and 10 more harnesses in Cloudroom with your own subscriptions and model accounts.
A harness is the coding agent that does the work. Cloudroom runs it and saves its history. The harness keeps its own agent loop, tools, and model calls.
Supported harnesses
- Claude Code: Anthropic's coding agent. Local and Cloud.
- Codex: OpenAI's coding agent. Local and Cloud.
- Pi Agent: an open-source coding agent that works with many model providers. Local and Cloud.
- OpenCode, fx, Hermes, Devin, Droid, and OpenClaw: Local and Cloud.
- Kimi Code: Moonshot's coding agent. Local after a one-click install in Settings → Providers, and Cloud.
- Agent Zero: an open-source agent that runs on your own Agent Zero server. Local after a one-click install of its
a0CLI in Settings → Providers, and Cloud. - Cursor and Grok: Local only.
Local threads run all 13 harnesses. Cloud threads run Claude Code, Codex, Pi Agent, OpenCode, fx, Hermes, Devin, Droid, Kimi Code, OpenClaw, and Agent Zero. Teleport works for Claude Code, Codex, Pi Agent, and OpenCode.
Who does what
flowchart TB
subgraph cloudroom [Cloudroom]
a[Starts and stops agents]
b[Queues your messages]
c[Saves and replays history]
d[Recovers after restarts]
end
subgraph harness [Harness]
e[Agent loop and tools]
f[Model calls]
g[Retries and compaction]
end
cloudroom --> harness
harness --> provider[(Your model provider)]Your subscriptions, no markup
Each harness signs in with your own account and calls its provider directly. Cloudroom never sits between your agent and the model, and never bills you for tokens.
Connect Claude Code
- Click Connect Claude. It appears next to the harness tabs in the model picker and in Settings → Providers.
- Local: sign in on your Mac and finish in the browser.
- Cloud: approve in the browser, then paste Claude's one-time code into the panel. Never paste it into the chat.
Connect Codex
After you sign in to Cloudroom, the app copies your existing Codex login from your Mac into the cloud and checks it. If there is none, Connect Codex opens OpenAI's official sign-in. You can reopen it in Settings → Machines → Agent logins → Codex.
Connect Pi Agent
Pi works with many providers. Pick a model as provider/model, for example openai-codex/gpt-6-astra. Sign in to that provider for Pi once. Pi's login is separate from Codex and Claude.
If Pi has no models yet, open the model picker and click Sign in to Pi. A terminal opens with Pi. Type /login, pick a provider, and sign in. Then open the picker again, and your models appear.
Use your own model server with Pi
Pi works with any server that speaks the OpenAI API, such as Ollama, LM Studio, vLLM, or a proxy. Add it to ~/.pi/agent/models.json on your Mac:
{
"providers": {
"ollama": {
"baseUrl": "http://localhost:11434/v1",
"api": "openai-completions",
"apiKey": "ollama",
"models": [{ "id": "qwen2.5-coder:7b" }]
}
}
}Then pick ollama/qwen2.5-coder:7b in the model picker. If it doesn't appear, run room-cli provider models pi --restart to reload Pi's models.
- Keys:
apiKeycan be the key itself,$NAMEfor an environment variable, or!commandto run a command, such as a password manager lookup. Pi's docs at pi.dev list every field. - Cloud threads: when you start a Cloud thread with Pi, Cloudroom copies your Mac's custom providers into its sandbox. This needs Settings → Cloud environment → Auth my agents in the cloud automatically turned on. Your Mac fills in
$NAMEand!commandkeys first and sends the result. If the cloud already has a provider with the same name, your Mac's version wins. - localhost:
localhostmeans the machine the thread runs on. A Local thread reaches a model server on your Mac. A Cloud thread can't, because itslocalhostis its own sandbox. For Cloud threads, use a server with a public URL.
Other harnesses
Install the harness's CLI on your Mac and sign in to it there. Local threads use that login.
Cloud Hermes uses your Mac's Hermes login, API keys, and model settings from ~/.hermes, copied automatically. It runs its commands in the cloud and sees Cloudroom's skills. Your Hermes memory stays on your Mac.
Cloud Devin uses your Mac's Devin login, copied automatically. Sign in with devin auth login on your Mac first.
Cloud Droid signs in with a Factory API key, not your Mac's login. Create one at app.factory.ai/settings/api-keys and add it as FACTORY_API_KEY under API keys in Settings → Cloud environment.
Cloud Kimi Code uses your computer's Kimi login and model settings from ~/.kimi-code, copied automatically. Sign in with kimi login on your computer first. Kimi renews its login on each machine, so your computer or the cloud may ask you to sign in again.
OpenClaw runs through its Gateway. Local threads use the Gateway on your Mac and start it if it's off. Cloud OpenClaw uses your Mac's OpenClaw sign-ins, API keys, and model settings, copied automatically. Claude plans are never copied: Anthropic allows them only in Claude Code. Your OpenClaw memory, persona, and chat apps stay on your Mac.
Agent Zero runs on your own Agent Zero server, usually in Docker. Its a0 CLI sends each message there, and the agent edits files and runs commands in the thread's folder through it. Local threads use the server a0 finds on your computer. Cloud threads need a server the cloud can reach, such as one on a VPS. Sign in to it with a0 on your computer and Cloudroom copies that login, or add AGENT_ZERO_HOST (and A0_USERNAME and A0_PASSWORD if it has a login) in Settings → Cloud environment. Your model is set on the Agent Zero server.
If a start fails
If a harness is not connected, Cloudroom keeps your first message. Connect the harness, then click Retry start. Nothing is lost, and running threads are not affected.
sequenceDiagram participant You participant App as Cloudroom participant H as Harness You->>App: Send a task App->>H: Start H-->>App: Not signed in App-->>You: Keeps your message, shows Connect You->>App: Connect, then Retry start App->>H: Start again H-->>You: Replies stream in