Six agent windows open, and you're the only thing connecting them.
That's how most engineers build with AI today. Keel is the environment I'm building to change it: one place where every agent you use connects, works alongside the others, and stays in view.
A harness for your harnesses.
Every coding agent comes inside its own harness: its own window, its own memory, its own rules. Keel doesn't replace them. You connect what you already use, a Claude Code subscription, Codex, Kimi Code, Cursor's agent, command-line agents, or your own model providers, and Keel becomes the place where you work with all of them.
It isn't built to produce a finished app in one shot. It's built for engineers who build every day and want the whole operation in front of them.
Keel
- Your subscriptions
- Your model providers
- Command‑line agents
- Editor agents
- Your own agents
- Open‑source tools
Every session, on one canvas.
In Keel, each session is a node on an open canvas. You can see what every agent is doing, group sessions by project, by milestone, or any way you think, and follow how work moves between them. When you come back from lunch, you don't read six scrollbacks. You look at the canvas.
The canvas
Group it your way
Agents that hand work to each other.
Sessions in Keel can talk to each other and delegate. You set up fleets, each with its own persona, skills, instructions, and rules: one for review, one for tests, one for research. When a session hits something outside its lane, it passes the work to the fleet built for it, and gets the result back.
A delegation system comes built in, and so does TeamOS, so a whole team works from one shared brain and one set of rules.
Fleets
Let them argue. Let them loop.
Some decisions are better when two sessions argue them out. Keel can set two sessions against each other on a question and show you both cases before you choose. For large tasks, you can set up a loop: do the work, check it, correct it, and repeat until a condition you chose is met. You describe the workflow the way you think about it, and Keel runs it.
A debate
Cache it: the data rarely changes.
Faster for every visitor.
Don't cache it: prices change hourly.
A stale price costs more than a slow page.
A loop
Round 3
Stop when: tests pass and the review is clean.
The things you do a hundred times a day.
Keel keeps track of tokens and cost for every session and every project, so you can see where the spend goes and what it bought. You can save a prompt and use it again. Memory is shared across sessions, and Keel learns from how you work and keeps what's worth keeping. And wherever a good open-source tool already solves one of these pains, I'd rather build it in than build it again.
Where the tokens went
Saved prompts
- Review this like a senior engineer
- Write tests before the code
- Summarize the day
As much control as you want, and no more.
Sometimes you want to talk an idea through, slowly, with every step explained. Sometimes you want to hand off a task and come back when it's finished. Sometimes you want a loop running all night. Keel lets you choose the level of control for each piece of work, and change it halfway through.
The control dial
Handed to the right fleet. You see the result.
The engineer should be steering, not copying and pasting.
Keel is for the engineers who work with agents every day. Plinth goes the other way: one run at a time, with every decision in your hands.
A shared brain and one set of rules for a team that builds with agents.
Previous: Caddy