Agents that build while you sleep.
An open source software factory for Claude Code, on machines you own.
fctl.sh · Features · Docs · Getting started · Self-hosting · Architecture
fctl runs Claude Code agents in sandboxes on your own Mac (or Linux box), around the clock, on your Claude subscription — never on API billing. Start work from the web app, the terminal or your phone; follow every session live; message an agent to steer it, with screenshots and files; share projects and sessions with your team.
curl -fsSL https://fctl.sh/install | sh
fctl login app.fctl.sh
fctl host install # this machine now runs agents
fctl run -p my-app "Add a dark mode toggle"Prefer Homebrew? brew install hyperstructure/tap/fctl installs the same
release.
Pointing an agent at fctl? llms.txt, also at https://fctl.sh/llms.txt, tells it what fctl is and how to drive it.
A runner is a machine someone on the team connected with one command: the
Mac mini in the cupboard, the Linux box under the desk, your own laptop.
fctl host install registers it and starts a service that survives a
reboot and updates itself. It takes the sessions of the workspaces it
serves, so your code is built and tested on hardware you own.
Each session gets a fresh Linux sandbox of its own: a lightweight VM on a Mac, a Firecracker microVM or a container on Linux. fctl clones the repository, runs the project's setup script and lets Claude Code work without stopping to ask for permission. Start from fctl's image, one from a registry, or a Dockerfile in the repository.
Everything the app does, the CLI does too: start a session and watch it
stream, steer it, search every transcript, manage projects, runners and
schedules. Add --json to any command, for scripts and for other agents
that hand work to fctl.
A workspace is the people who work together, with their projects, runners and GitHub accounts; roles go from viewer to owner. A project is a repository and everything its agents need: the model they start with, instructions, secrets, and the most they may do on GitHub.
Turn them on, and an agent reviews each new commit of every pull request in a project's repositories, whoever wrote it, with findings on the lines they are about. Findings on an fctl agent's pull request go back to that agent, which fixes what's real and answers in the thread.
Delegate a Linear issue to fctl, or mention @fctl in a comment, and a
session starts in the project that takes the issue's team, as you, on your
Claude plan. The agent answers in the ticket's thread, and the pull request
comes back linked. Projects that track their work in GitHub issues
instead have their agents read and write tickets there.
Agents leave notes for the agents after them: the command that works, the trap they fell into, how the tests really behave. Every session starts with the project's notes, and when an agent finishes a piece of work it records how it went, for the people who follow along.
Give an agent a task and a time, once or again and again: dependency updates every Monday, triage every morning, benchmarks every night. Each run is a session of its own, on the subscription of whoever set it up, and other tools can start one with an access token.
Read about more features on fctl.sh, and everything else in the documentation, whose source is in docs/.
You start and follow work from the web app on your desk or phone, the
fctl CLI in a terminal, script or CI job, or the Claude apps through
Remote Control. They all talk to the control plane: a Rust Worker compiled
to wasm, with Durable Objects and D1, that runs on Cloudflare or in one
container anywhere. It keeps every session's events and hands sessions to
your machines.
Each machine runs fctl host, which dials out to the control plane, so it
needs no open ports. A Mac runs every session in a VM of its own with
Apple's container; a Linux box, in a Firecracker microVM or a Docker
container. In each sandbox, fctl-agent runs Claude Code and journals its
events, and the agent works on a branch, pushes it and opens a pull
request.
- Subscriptions only. API keys are stripped from every sandbox, sessions must report a subscription credential, and a session pauses before it would use paid extra usage.
- Nothing is lost. Agents journal every event until the control plane acknowledges it; links reconnect and replay; hosts recover running sessions after a restart.
- Your tracker. Each project chooses where its agents read and write tickets: GitHub Issues, Linear, or none at all.
- Follow along anywhere. With a full Claude login, sessions join Remote Control and show up in the Claude apps.
- Open source and yours to run. All of fctl is MIT licensed. Use the
managed control plane at app.fctl.sh, or host your own: on your
Cloudflare account with
cargo xtask deploy, or without one, on a spare machine or a Kubernetes cluster, withdocker run ghcr.io/hyperstructure/fctl-control(self-hosting).
The control plane, API and web app together, runs in one container with its state in a volume, so trying fctl on your own machine needs only Docker. Start it, open the setup link its log prints to create the first account, then point the CLI at it instead of app.fctl.sh:
docker run -d --name fctl --restart unless-stopped \
-p 8787:8787 -v fctl-data:/data \
-e FCTL_PUBLIC_URL=http://localhost:8787 \
ghcr.io/hyperstructure/fctl-control
docker logs fctl # the link that creates the first account
fctl login http://localhost:8787
fctl host install # this machine runs its agentsSelf-hosting has the full details: Kubernetes and Cloudflare, HTTPS, settings, GitHub, Linear, sign-in, email and backups.
cargo xtask dev # control plane, web app and a host running fake-claude
cargo xtask check # what CI checksStart with AGENTS.md (written for coding agents, useful for people) and docs/development.md. The plan is in ROADMAP.md. The design system lives in hyperstructure/design and fctl.sh in hyperstructure/fctl-site.