Every agent. One continuous take.
OpenAgents is an open-source harness for AI agents. Bring your own API keys or plug in the agent CLIs you already use, and run Claude, Codex, Gemini and local models side by side in one workspace. When a session hits its limit, the live task moves to the next agent with its full context.
In development. Open source. Early access soon.
Agents got their own computers. You lost the keys.
Since August, xAI, Meta and OpenAI have each shipped an always-on agent with its own cloud computer. They are polished. Each one lives in its vendor's cloud, on its vendor's plan, with the models its vendor picks.
The open agents, OpenClaw and Hermes Agent, hand the keys back: your machine, your models, MIT-licensed code. You mostly drive them from a terminal or a chat thread.
Nobody has put the two together: open and model-agnostic, with an interface worth working in all day.
Hit the limit. Keep the take.
Usage limits land at the worst moment. OpenAgents keeps a continuity report while the agent works, so the next agent never starts from zero.
- Written as it works
- Goal, decisions, files touched, commands run and the next step, updated after every step.
- Handed over in one cut
- When a session ends, the next agent you have connected, from any provider, reads the report and continues.
- A file you can read
- The report is designed as plain Markdown in your project. Open it, edit it, commit it.
A real workspace, not another terminal tab.
Four agents, four angles, one screen. Watch each one work, cut to the one that needs you, and step in without breaking its flow.
- Every agent's plan, output, diffs and browser in one view.
- Desktop and web, sharing the same state.
- Keyboard first. Every action has a shortcut and a menu path.
Cast the right model for every role.
Connect models three ways: your own API keys, the agent CLIs you already have installed and signed in (Claude Code, Codex CLI, Gemini CLI), or a model running on your own machine. Give the planner, the builder and the reviewer to different models, and swap one without touching the others.
Memory you can read, correct and delete.
What an agent learns about you and your project is kept as plain files you own. Read every entry. Fix the wrong ones. Strike out what it should forget.
Rolling while you sleep.
Schedule recurring work, keep long jobs running on your own machine or server, and get a message when a take needs your call.
Wired into the tools you already use.
MCP servers, Git, a browser, your terminal and your files. Reach your agents from Slack, Discord or Telegram. Connect a tool once and every agent can use it.
Nothing ships without your sign-off.
Decide what an agent may do on its own and what it has to ask about first: pushing to main, sending a message, spending money, deleting files. Try it on the requests below.
-
Builder · CodexPush 3 commits to mainBranch retry-worker, CI passedWaiting for you
-
Researcher · GeminiSend a message to #releaseDraft of the release notesWaiting for you
-
Planner · ClaudeSpend $4.20 on API callsOver today's limit of $3.00Waiting for you
-
Builder · CodexDelete 14 files in /legacyNo references foundWaiting for you
Where OpenAgents fits.
What the agents people talk about this autumn are built around, from their own public descriptions.
| Agent | Code | Models | Runs on | Main interface |
|---|---|---|---|---|
| OpenAgentsIn development | Open source | Any provider: your API keys, your installed agent CLIs, local models | Your machine or your server | Desktop and web workspace, plus chat apps |
| OpenAI DotsSource: 9to5Google, 29 Sep 2026 (opens in a new tab) | Closed | GPT-6 Astra | Its own cloud computer | ChatGPT, Slack, Teams |
| xAI Grok BotSource: Keywords Everywhere, Sep 2026 (opens in a new tab) | Closed | Not stated | Its own cloud computer | SuperGrok plans, X connector |
| Meta MuseSource: iClarified, 8 Sep 2026 (opens in a new tab) | Closed | Muse Spark | Muse Secure VM, Meta's cloud | Web, iOS and Android apps, WhatsApp |
| OpenClawSource: GitHub, Oct 2026 (opens in a new tab) | Open source, MIT | Hosted and local providers | Your machine or server | Chat apps, CLI, TUI, web control panel |
| Hermes AgentSource: GitHub, Oct 2026 (opens in a new tab) | Open source, MIT | Many providers and custom endpoints | Your machine, a VPS or serverless | CLI and TUI, chat apps |
Descriptions of other products are summaries of the public sources linked in each row, as of October 2026. Product names belong to their owners. OpenAgents is not affiliated with any of them.
Open, all the way down.
- Open source
- The harness, the workspace and the report format, in public from the first commit.
- Bring your own models
- Subscriptions, API keys or local weights. No markup on tokens, because we never resell them.
- Local first
- Reports, memory and settings are designed to live on your disk as files you can read.
- Asks before it acts
- By default, anything that leaves your machine or cannot be undone will wait for your approval.
The shooting schedule.
No dates until they are real. The order is decided; early-access testers decide what moves up.
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
Be in the first screen test.
Early access goes to a small group first. Testers shape what ships in the first release.
We store your email, the time you signed up and anything you add above, only to send your early-access invite and get in touch. No newsletter, no tracking, and we never share it.
Questions
Is OpenAgents free?
The code will be open source. Models cost whatever your provider charges on your own subscription or API key, and OpenAgents is not designed to add anything on top.
Is moving a task between subscriptions allowed?
OpenAgents moves your own work between accounts you own. It does not share accounts or get around limits. Check each provider's terms for how you use their service.
How does OpenAgents connect to models?
Three ways, mixed freely: API keys you paste in (stored on your machine, never sent to us), agent CLIs you already have installed and signed in, such as Claude Code, Codex CLI or Gemini CLI, which OpenAgents drives as harnesses, and local models through Ollama-compatible servers.
Which providers come first?
Claude, Codex, Gemini and local models through Ollama-compatible servers are the first targets. Testers decide what comes next.
Where does my data live?
On your machine. OpenAgents is designed to keep continuity reports, memory and settings as plain files in your project or home folder, and to send prompts only to the model providers you connect.
What license will it use?
An OSI-approved open-source license, chosen before the first public release.
When can I try it?
Early access opens to the first testers before the public release. No date is set yet.
Tell us what to build.
Ideas, questions, a workflow that keeps breaking, an agent you want supported. We read every message. Nothing you send is published.