Self-hosted coding agents, in your own cloud.

A coding agent is not a chat window. It reads your code, runs commands, installs packages, holds keys to your repositories and pushes changes. Where it does that work is a real decision: on a developer’s laptop, on its vendor’s computers, or on computers your company controls. Self-hosting is the third choice.

Claude Code’s cloud sessions, for example, run on Anthropic’s computers unless your company sets up its own.¹ That is the easy path, and for many teams the right one. Teams self-host when the work has to sit on their side: the checkouts, the build artifacts, the keys an agent uses and the files it makes. Some need their agents to reach internal services. Some answer to a security team that asks where the code lives. Some want one bill they can read themselves.

One thing self-hosting does not change is the model. Claude and GPT still run at Anthropic and OpenAI, and what an agent sends the model, your code included, goes there under your plan’s terms. Self-hosting moves where the agent works, not where the model thinks.

Three ways come up again and again. Each is a fair answer to a different question.

Whichever way you choose, the agent is the small part. To take a job from one sentence to something live, it needs a computer that keeps its files between turns, a repository to push to, checks that run on every push, and a person’s yes before protected code lands. It needs somewhere to publish what it built, a way to use your keys without seeing them in a chat, and the notes and decisions your team already wrote down, so it does not start from zero every morning. Often it needs an email address too.

On a bare server, each of those is a project. Together they are the difference between an agent that suggests a change and one that finishes the job.

Every piece on that list can be built, and teams with strong engineers often start to. Count what it adds up to. A computer for each agent that keeps its files, sleeps when idle and wakes when work arrives. A git host with protected branches, checks on every push, and a proposal a person can approve from a phone. Somewhere to publish a site on your own domain, with its certificates. A store for keys that hands an agent the value without showing it in a chat. Notes and pages the agents read before a job and write back to after it. Mail addresses that send and receive. And one place where people talk to all of it.

Then it has to keep working. Claude Code, Codex and OpenCode change under you, with new versions, new settings and new ways to sign in, and every agent’s computer has to follow without losing its work. Keys have to stay encrypted and out of transcripts, and an agent must not be able to merge its own change or widen its own access. The disks need backups you have actually restored, and an upgrade needs a way back when it fails. Someone on your team owns all of it, including when it breaks at night.

We know the size because we built it. The open-source repository holds more than 230,000 lines of code, with close to 400 test files beside them, and a whole document on how an update reaches agents’ computers that are already running. It is a large system to build, and a larger one to keep.

That is the real cost of building it yourself: not the first version, but the attention after it. A company’s attention belongs on its own product or service. A workplace for agents is something your agents need, not something your customers buy.

Building it yourself is still the right call in some cases. If agent infrastructure is your product, build it. If your agents must live inside a network or a cloud you cannot leave, or you need controls this does not have, such as single sign-on, you may have no other choice. If you need one agent for one job, such as a review in the CI you already run, a step in that CI is smaller than any platform. And if you want most of this but not all of it, the code is open source under the AGPL-3.0: start from ours and change what you need.

Your own Agentic Cloud is all of that, installed in one Fly organization of yours, on your own domain, by one command. Each agent gets a computer of its own that sleeps after about three quiet minutes and wakes in seconds when there is work. Claude Code, Codex and OpenCode work side by side, each on your people’s own Claude, ChatGPT or OpenCode Go plan. Around them sit git with CI/CD, sites on your domain, Drive, notes, knowledge pages and, once an admin turns it on, email. A change to protected code waits for a person’s yes, or merges by itself once the owner turns on auto-merge, and the owner is told either way.

None of your data comes to us: only a daily report of counts and the contact you chose, which your admin sees and can turn off. Our clean run on 29 October took 13 min 24 s from a new Fly organization to ready. An admin upgrades from the Workspace, and an upgrade that fails puts the old release back by itself.³ It is free for now, with no licence and no key, and the whole platform is open source under the AGPL-3.0, so your team can read what it creates before anyone says yes.

Know these before you install. It runs on Fly, not on AWS or Google Cloud, so credits with those clouds do not apply; why it runs on Fly is its own page. The installer does not join it to your company’s existing network, so agents that must reach an internal database need a way in that you set up. There is no single sign-on: people sign in with a name and a password, and an authenticator code if they choose.

The models are Anthropic’s and OpenAI’s, reached through your own plans, so this is not a way to keep code away from a model provider. And if one developer wants one agent on one laptop, Claude Code on its own is simpler.