Meet Archestra 1.4 "Arpa": Remote Agents
Archestra 1.4 "Arpa" is a big release. Here we focus on its headliner: a self-hosted runtime for remote coding agents, with handoff from your laptop to the cloud.
2026-10-21 · Matvey Kukuy, CEO
Hey everyone! Today we're shipping the long-awaited Archestra 1.4, codename "Arpa" 🎉
"Arpa" is a nod to OpenAPPA, which also ships in this release. It will get its own blog post soon. Follow us on LinkedIn, X, and YouTube, or join our Slack community, so you don't miss it.
As always, let's start with some numbers. This release packs 712 bug fixes and over 1,200 changes, all of which you can check out on GitHub. Seventeen developers worked on it, spending about 5,600 hours and roughly 250 billion tokens. Not that we're bragging; we counted, so we might as well share.
Remote agents
Remote agents go by many names: "meta-harness," "cloud agents," or a product name like Claude Code on the web at Anthropic, Codex cloud at OpenAI, or Orbs at Amp.
Archestra is vendor-neutral and open source, and we wanted remote agents to stay that way, so we are NOT announcing a proprietary coding agent. You already have a favorite one, and now our cloud runtime can run it.
Out of the box, Archestra supports Claude Code, Codex, OpenCode, Hermes, and OpenClaw. Best of all, it lets you use your subscription tokens. And yes, as usual, it runs entirely in your infrastructure. No opaque VMs in an AI vendor's cloud, no lock-in to any one of them. Full control over the deployment in your Kubernetes cluster. In other words, if the Anthropic or OpenAI sales team is already circling your leadership, pitching cloud agents as a reason to lock you in, you know what to do.
Naturally, remote agents work hand in hand with everything already in the platform:
- They get their tools through our MCP Gateway.
- Logs are collected in one central place.
- They can't leak data, because they run behind guardrails (more on that soon).
There's a lot of room for customization. For example, you can build your own Docker image with the Pi agent and run it in our agent runtime. And by the way, these agents become reachable over A2A and the API.
Yes, you can now call Claude Code through an API. If "feature factory" just crossed your mind, we'll write about that soon.
Handing off from a local agent to a remote one
In Archestra 1.3 we introduced the Connect Page, a self-serve page that lets anyone in the company connect their agents to Archestra in seconds. In this release we've reworked it substantially.
The Connect Page now also configures your agents to hand work off to a cloud agent. For example, you can start a task in Claude Code, say "finish this in Archestra," and Claude will spin up Codex and carefully pass it the full context of the task.
Admins can also inject a system prompt into connected agents, so you can nudge coding agents to push heavy tasks to the cloud without pulling developers away from their work.
That's it for today. Install Archestra, tell us what you think, and join our community!
Soon we'll write about applying guardrails on the LLM proxy level, so follow us on LinkedIn, X, and YouTube, or join our Slack community.