One protocol, not per-agent glue
Codex and Claude Code run on hosted boxes today. Both speak ACP, so the same box could drive any of the 30+ registry agents — the protocol is the integration, not code we write per agent.
Built on ACP — the open agent protocol
A cloud box for
your coding agent.
Driven from anywhere.
OpenAgent runs Codex or Claude Code on a dedicated cloud box and puts it on a secure line you can pick up from any device. Bring your own OpenAI or Anthropic key — start a task at your desk, approve a permission from your phone, come back hours later to a session that never dropped.
No SSH tunnels · Billed in machine-time, never tokens · Your API key goes to your box, not our models
you · iPhonerun the test suite and fix whatever breaks
codexRunning npm test…
codex3 failures in auth.spec.ts — patching token expiry check
Permission — codex wants to run npm test
Allow onceAlwaysDeny
codex214 passing. Branch fix/token-expiry ready for review.
— reconnected after 2h away · 0 events lost —
Built on ACP — the open protocol 30+ agents speak
- Codex
- Claude Code
- Gemini CLI
- Goose
- OpenCode
- + the whole ACP registry
Codex and Claude Code are live on hosted boxes today. Because OpenAgent targets the protocol, the rest of the registry is what it opens up — not per-agent work we have to ship.
How it works
One protocol, carried across the network
We didn't invent a session protocol. ACP — JSON-RPC between client and agent, the LSP of coding agents — is the spine. OpenAgent carries it over the wire, from a box in the cloud to whatever device is in your hand.
-
Your agent already speaks ACP
An open standard from Zed and JetBrains, spoken by 30+ agents and counting. OpenAgent targets the protocol, not a vendor CLI — so the same box runs Codex or Claude Code today, and every future registry agent comes for free.
-
The box dials out to the relay
We boot a dedicated cloud box for your agent. A small sidecar on it opens a TLS WebSocket out to our relay — no inbound ports, no SSH tunnel. The agent's lifetime is decoupled from your connection, so a dropped socket never kills a running task.
-
Any device picks up the line
Phone, laptop, browser — open the web client and connect through the relay. Reconnect replays exactly what you missed, and permission prompts you weren't there for are re-delivered, not dropped.
Features
Built for sessions that outlive the connection
Permission prompts follow you
Agents block on approval. OpenAgent holds every unanswered request and re-delivers it wherever you reconnect — a locked phone screen misses nothing.
Sessions survive everything
Every event is sequence-numbered and buffered. Come back after hours and replay the exact stream you missed, mid-task, mid-turn.
Bring your own key
Your OpenAI or Anthropic key is delivered in-band to your box only, then kept on the box's own disk. Model traffic never touches our servers — zero token margin, by construction.
Isolated by default
Every box is a per-user micro-VM. Kernel egress rules block the cloud metadata endpoint and cross-tenant traffic in every mode; lock a box down further with a per-template network policy.
Minimal surface
We build only the missing piece and consume the rest — TLS is Caddy, supervision is systemd, framing is the official ACP SDK. Owned code measures in hundreds of lines, not thousands.
The platform
The tool stands alone.
The platform grows on top.
Because the core exposes raw ACP, everything above it is just another client — accounts, APIs, hosted machines. All optional, none of it changes the core.
Live
Hosted machines
We run a dedicated box for your agent: persistent environment, your choice of Codex or Claude Code, your credential — billed in machine-time, never tokens.
Live
One API, every agent
An OpenAI-shaped /v1/chat/completions endpoint that drives whichever agent a user connects. Developers get the familiar SDK experience; users pick the agent and bring the credential.
Live
Dev apps
Third-party apps act on a user's behalf through OAuth consent — scoped, revocable, enforced at the relay. The core never even learns apps exist.
Running the core yourself, on your own machine, is on the roadmap — closed for now while the hosted product finds its feet.
No agent vendor will neutrally host their rival's agent on the user's own terms. We will.
Zero token margin, structurally. We never touch inference billing. What you pay us is for machines and developer tooling, never a markup on tokens.
No lock-in
Raw ACP is the API
There's no proprietary session format underneath. Any spec-compliant ACP client — ours, Zed's, one you write — connects directly. If OpenAgent vanished tomorrow, your setup would still speak an open, Apache-licensed protocol.
→ initialize
← initialize (agent capabilities)
→ session/new
→ session/prompt "fix the flaky test"
← session/update × n, streamed
← session/request_permission ⟵ carried to your phoneFAQ
Fair questions
Does my code or key ever pass through your servers?
Your box is a per-user micro-VM only you can reach, gated by a token. Your OpenAI or Anthropic key is delivered in-band to that box and stays on its own disk; model traffic goes straight from your box to the provider, so we never see or meter your inference. Your code lives on the box's disk — destroy the box and it's gone.
Which agents can I run?
At launch, Codex (on your OpenAI key) and Claude Code (on your Anthropic key). Both are ACP agents, so widening to more of the 30+ agent registry is a protocol away, not a rewrite.
How is this different from wrapping one agent's app?
Single-vendor remote tools wrap one CLI and break when it changes. OpenAgent targets the protocol layer those agents already standardize on, so one codebase drives all of them — and the box, session-resume, and permission re-delivery work the same whichever agent you pick.
What does it cost?
You prepay a credit pack and we meter the box by the second while it runs — 4.32¢/hour for a 1 GB box, nothing while it's suspended. No subscription, and never a markup on tokens; you pay your model provider directly. See pricing for the full rate card.
Is it open source?
The protocol is: ACP is an open, Apache-licensed standard, and raw ACP over wss is our API — no lock-in. The core tool's own source is closed for now while we launch the hosted product; that may change.
When can I use it?
The hosted product is live end to end — a stranger can sign up, buy a pack, spin up a box, and drive an agent from their phone. We're opening it in waves; leave your email to get early access.
Your agents shouldn't need you
in the room.
A cloud box, your agent, your key, one connection. Be first in line when early access opens.
No spam — one email when it's ready.