Step in with a real editor
Every workspace ships VS Code in the browser and a real shell, one click from the chat. Edit a file by hand, run a command, then hand it back to the agent.
Self-hosted. Built on opencode.
The always-on
AI coding platform.
Start a task at your desk, close the laptop, and pick up the same live session from your phone.
- Runs anywhere Docker runs
- Any model, cloud or local
- One-command install
Always on
The session never stops.
- 6:12 pm At your desk Kick off a refactor across the repo, then close the tab.
- 6:38 pm On your phone A push says the tests pass and the agent wants a review.
- 6:39 pm From the couch You approve, and it keeps going. The session never dropped.
When the agent hits a question or a permission gate, it pushes to your phone. You answer from anywhere, and the run picks up where it left off.
The workspace
The agent does the work. You stay in control.
Every session runs in a real dev environment. Watch the agent test in a live browser, check any diff with the built-in differ, and step into a full editor or terminal whenever you want.
Watch it drive a browser
A live screencast slides in so you can see the agent test its own work, and the same dev server opens in your tab with one click.
Review with the built-in differ
Every file the agent touches is one click away as a syntax-highlighted diff, with a branch strip and live line counts, so you can check the work when you want to.
Remote execution
Run it where the work belongs.
Register a machine once, then send any session to run on it. Check the status of each environment, spread work across containers, or push a heavy build to a bigger box.
Secrets and integrations
Credentials it never sees.
Manage OAuth integrations and a secrets vault from the dashboard. When the agent needs a credential, the broker injects it at the network edge and never exposes the value.
Custom agents
Design your own agents.
Build an orchestrator that plans and subagents that execute, each with its own model and permissions. Hand the busywork to a fast model and the decisions to a strong one.
Orchestrator plans the work strong model
Builder writes the code fast model
Reviewer checks the diff fast model
Everything else
More than a coding agent.
- Automations YAML workflows with human approval gates.
- Secrets broker A vault plus injection at the network edge.
- Voice input Dictate with any Whisper-compatible endpoint.
- Durable memory Files the agent always keeps in mind.
- Custom agents Orchestrators and subagents with per-agent models.
- Installable PWA Add it to your phone and get push notifications.
Works with your stack
Bring your own models.
Point Kellix at a hosted API or a model running on your own hardware. It runs anywhere Docker runs, so the same instance works on a mini PC, a workstation, or a cloud box.
- OpenAI
- Anthropic
- Ollama
- vLLM
- Any OpenAI-compatible API
Install
One command, then you are in.
It runs anywhere Docker runs. One install, and everyone on the instance gets their own isolated workspace.
curl -fsSL https://downloads.kellix.ai/main/install.sh | bash- 1
Install
Run the command on any machine with Docker.
- 2
Open the setup URL
Create your admin profile and connect a provider.
- 3
Start building
Reach your live sessions from any device.
Questions
Common questions.
What hardware do I need?
Anything that runs Docker. A mini PC or an old laptop is plenty, and heavier local models simply want more RAM or a GPU.
Which models can I use?
Any OpenAI-compatible API, or a local runtime like Ollama or vLLM. You bring the keys, and you can pair a strong model for planning with a fast one for the busywork.
How do updates work?
One command: kellix update. It pulls the new image and restarts in place, so your sessions and settings carry over.
Where does my data live?
On the machine you installed it on, and nowhere else. Encrypted backups are built in, so you can move or restore an instance safely.
Can multiple people use it?
Yes. One install serves everyone, and each person gets an isolated, private workspace with their own sessions and secrets.
Is my code sent anywhere?
Only to the model provider you pick, and nothing else leaves the box. Point it at a local model and nothing leaves at all.