A desktop workspace for Claude Code. Every task gets its own git worktree, so several agents can be mid-edit on the same repository without touching each other's files.
The actual application, running here on sample data, not a video and not a mock. Pick a chapter below, or click in and drive it yourself.
Every chapter switches GROUP: a set of tabs you live in. Click in to drive it yourself.
Counted from the public repository when this page was built. First commit 11 February 2026.Check them yourself.
There is no new workflow to learn. The agents are the ones you already run, in a terminal, on your own account.
One conversation, and the things it needs: a terminal, a branch, a browser, the files it touched. Close the app and it comes back the same, scrollback included.
Split any pane, drag tabs between groups, keep three whole projects open beside each other. The layout persists, and syncs to your other machine.
Write a task, an agent takes it in a git worktree of its own, and hands back a commit and evidence. You read the diff and decide.
Runs the agents you already run
Claude Code · Codex · OpenCode · Gemini CLI · Anthropic API · OpenAI GPT · MCP · any ACP agent
A worktree per agent is not ours alone any more: Devin Desktop sells the same idea. What is still unusual is the board: the agent picks the card up itself, works it, and brings it back with evidence.
What should happen, where to look, how you will know it worked. Write it vaguely and it comes back vague.
It gets an isolated worktree on its own branch. Four agents can be mid-edit at the same time and none of them collide.
Committed work, a short summary, and a screenshot or clip of the thing running. Then it moves its own card to review.
Read the diff and approve. Landing it into main is a second click, and pushing is a third. None of them happen on their own.
Launched from a plain shell, Claude Code falls back to low effort. Topics passes --effort on every spawn, so a session opens at the top tier. Drop it per topic when the job is small and you would rather not pay for the deep thinking.
When an agent spawns a fleet of its own, the children appear in the session tree instead of disappearing into a pipe. You can see what each is doing and what it is spending while it does it.
Approving accepts the work. Landing merges it locally. Publishing pushes. Three decisions, three clicks, because "done" should mean you said so.
Each topic carries its own project, terminals and browser, and gets them back after a restart.
01 Organize
02 Run
dev or test from the sidebar and keep their output.
03 See
04 Ship
05 Reach
All of this is in the free app and works over a tunnel you run today. There is no hosted version and none is running. See the limits.
06 Own
Not features, which move in a release. Where the work runs, and who else can read it.
| Question | Topics | Conductor | Devin Desktop | Zed |
|---|---|---|---|---|
| Where the work runs | Your machine. No server of ours in the path. | Local, or their cloud, where chat messages are stored on their servers. | "Fleets of local and cloud agents." | Your machine. |
| Price | $0 for the whole app | Free $0 · Pro $50/mo · Teams $60/mo/user | Not stated on the product page | Personal $0 forever · Pro $10/mo · Business $30/seat/mo |
| Licence | MIT, open source | Not stated | Not stated | Open source |
| Account required | None | Required for the paid plans | Sign-up required | Not for the free plan |
Every cell about a product other than ours is quoted from their own public page, read 4 August 2026. The full table, six tools, and when each of them is the better answer →
Written by the people who make it, which is the only reason it is worth reading.
Start a session, watch it, finish. A workspace is overhead. Everything here is in service of the case where three or four are running and you have lost track of which one is waiting on you.
It will not. The agents run on your machine, so they stop when it sleeps. If that is the requirement, a product with cloud workspaces is the right shape and we say which ones on the comparison page.
Topics is not an IDE and is not becoming one. It hosts the terminal agents and the surfaces around them; you keep whatever editor you already use, open beside it or inside a pane.
There is no authentication layer at all. Not a weak one: none. Boards sync between your machines. Exposing one over a tunnel means putting your own auth in front of it.
The build is signed ad-hoc, so Gatekeeper stops it on first launch and since Sequoia the control-click bypass is gone. The three steps are below, and if that is a dealbreaker it is a fair one.
First commit 11 February 2026. Windows and Linux are built and released every time, and are used far less than macOS, which is exactly where you should expect to find the rough edges.
All of it shipped, all of it free, none of it worth its own headline.
@ mentions for files and topics.Cursor and Warp rebuilt the editor. Topics did not. It is a home for the terminal agents you already trust, where you can see what they are doing and step in when they need you. You keep the workflow you have. It gets a window worth looking at.
Everything that runs on your machine is free under MIT, permanently: every agent, every terminal, the whole board. What you can pay for is support, an invoice, and a say in what gets built, never a feature held back from the free app.
€0the whole product
Not a trial and not a tier. This is the app.
MIT. No sign-up, no email, nothing to cancel.
€5/ month, or €49 a year
You use it daily and would rather it kept being made.
Cancel whenever. Nothing in the app changes either way.
€99once, not a subscription
Everything Supporter gives, plus the hosted side when it exists.
Nothing hosted exists today. You are buying a promise with a refund attached to it, and you should read that as exactly what it is.
€490/ year, per company
Your finance department needs a counterparty, not a GitHub repo.
Unlimited seats. There is no per-seat count to police because there are no accounts.
No, and the shape of the product is the answer rather than the promise. Your data is SQLite and files on your own disk, the licence is MIT, and there is no account to close. Stop paying, or delete this page from your history, and the app carries on exactly as before.
Never. Topics talks to Anthropic or OpenAI with your key, or through the CLI you are already signed in to, and that bill stays between you and them. We do not resell inference or take a cut of it.
It cannot without a fork appearing the same week, which is what MIT is for. Nothing in the paid tiers is a capability withheld from the app. Read the lists again and notice that not one of them is a feature.
Then do it yourself, and here is how. Remote access is a cloudflared or tailscale tunnel with your own auth in front. Push notifications are ntfy. Off-machine backup is rclone or restic pointed at ~/.topics. All of it works today with the free app, and saying so out loud is the only thing that makes the sentence above it believable.
Every release, and what changed in it. Full changelog →
Free, open source, ready to run.
Topics is signed ad-hoc rather than with a paid Apple Developer certificate, so the first launch is blocked. Since macOS Sequoia the old control-click trick no longer works: Settings is the only route, and it takes about fifteen seconds:
You only do it once. If you would rather check first, the build is reproducible from the tag and every asset has a .sig beside it on the releases page.
There is a machine-readable page at /agents.md with the install steps, the ports and the MCP surface. Paste this into any coding agent:
Read https://topics.armonia.io/agents.md and install Topics on this machineEvery asset and checksum lives on the releases page.