Every agent. One continuous take.
OpenAgents: the open-source harness for AI agents.
In development. Open source. Early access soon.
Agents got their own computers.
Since August, xAI, Meta and OpenAI have each shipped an always-on agent that lives in its vendor's cloud, on its vendor's plan. Polished, and closed.
The open ones live in a terminal.
OpenClaw and Hermes Agent give you your machine, your models and MIT-licensed code. You drive them from a command line or a chat thread.
And every agent stops at the limit.
A usage cap ends the session halfway through the task. The next agent starts from zero, and you explain everything again.
OpenAgents keeps the take rolling.
One workspace for agents from every provider. When a session ends, the task moves to the next agent with its full context.
The continuity report.
Goal, decisions, files, commands and the next step, written as the agent works. A plain file you can read and commit.
- Goal
- Move billing webhooks to the new queue without losing events.
- Decided
- Keep the old endpoint live for 14 days. Idempotency key = event id.
- Touched
- src/billing/webhooks.ts · src/queue/worker.ts · 2 tests
- Ran
- npm test billing (41 passed)
- Next
- Add retry with backoff to worker.ts, then rerun the suite.
A workspace worth living in.
Four agents, four angles, one screen. Desktop and web, keyboard first.
- read issue #212
- split into 4 tasks
- assign: builder, tests
- watch: CI on branch
- update plan v3
- read issue #212
- split into 4 tasks
- assign: builder, tests
- watch: CI on branch
- update plan v3
- edit worker.ts +38 -12
- add backoff(5, 2x)
- run unit tests
- 41 passed, 0 failed
- commit: retry worker
- edit worker.ts +38 -12
- add backoff(5, 2x)
- run unit tests
- 41 passed, 0 failed
- commit: retry worker
- open staging dashboard
- trigger test webhook
- event 9f2c received
- screenshot saved
- compare with prod
- open staging dashboard
- trigger test webhook
- event 9f2c received
- screenshot saved
- compare with prod
- diff 3 files
- flag: missing timeout
- suggest: 10s ceiling
- waiting on you
- approve or edit?
- diff 3 files
- flag: missing timeout
- suggest: 10s ceiling
- waiting on you
- approve or edit?
Any model, any role.
Your subscriptions, your API keys, your local models. Swap one role without touching the rest.
| Role | Model | Billed through |
|---|---|---|
| Planner | Claude | Subscription |
| Builder | Codex | Subscription |
| Reviewer | Qwen, local | Your machine |
| Researcher | Gemini | API key |
Open and polished, at once.
Closed agents have the interface. Open agents have the freedom. OpenAgents is built to have both.
| Agent | Code | Main interface |
|---|---|---|
| OpenAgentsIn development | Open source | Desktop and web workspace, plus chat apps |
| OpenAI Dots | Closed | ChatGPT, Slack, Teams |
| xAI Grok Bot | Closed | SuperGrok plans, X connector |
| Meta Muse | Closed | Web, iOS and Android apps, WhatsApp |
| OpenClaw | Open source, MIT | Chat apps, CLI, TUI, web control panel |
| Hermes Agent | Open source, MIT | CLI and TUI, chat apps |
What we are shooting first.
Handoff and sessions across providers now. Memory, approvals and schedules next.
Now
- Continuity report and handoff between agents
- Sessions across Claude, Codex, Gemini and local models
- Desktop workspace
Next
- Memory editor
- Approval rules
- Scheduled and background runs
Later
- Chat app channels
- Shared team workspaces
- Plugin and skill directory
- In production
- Next
- Later
Join the first screen test.
We are looking for early-access testers who run agents every day, and contributors who want to build the harness with us.
Website