Draw the crew.
Run it on your own Claude.
Think Dify or n8n, but for agents that read and change code — on your own machine, with a verifier that can say no.
Concentus is a desktop application for wiring multi-agent flows on a canvas and running them against your own repositories, your own databases and the Claude subscription already sitting on your machine. No server to deploy, no database to provision, no API key to paste.
The flow you draw is the flow that runs.
Drop agents, tools and data sources onto a canvas and connect them. A coordinator hands work to a crew; the crew hands its results to a verifier. There is no YAML hiding behind the picture and no orchestration code to write — the graph compiles straight into a multi-agent session.
v7 · 5 nodes · 4 edges
Nodes, edges, and the ports that say what may connect to what.
A multi-agent session, launched and streamed over a socket.
Commands routed to a session that is already running.
Every run remembers the flow revision it executed.
Sixteen kinds of node. One vocabulary.
Each is a real card on the canvas: an icon, a name, a badge, and a coloured spine that says what it is from across the room. Chains, not a flat roster — an agent can feed the next, branch on a condition, fan out over a list, or call another flow entirely.
Plans the work and hands it out.
Does one job, in its own context.
Any MCP server, by URL.
A query's result, handed in as context.
A repository the crew can read.
Your documents, retrieved when relevant.
Any HTTP endpoint, by typing its URL.
Checks the work before it counts as done.
A real if/else — no input can fall between the branches. Every block has an on-error output too.
Fan the same work across a list.
Where parallel branches come back together.
One flow calling another.
Where a run starts: typed, scheduled, a webhook, or an incoming mail.
Sends the result by SMTP when the run gets there.
A note on the canvas. Group blocks to move them together.
Watch it work. Take the wheel.
A run streams as it happens: every agent's own output, filterable one agent at a time, each holding its colour across the whole transcript. Ask an agent to approve its plan before it acts, and nothing touches your disk until you say so.
What the adversarial verifier killed, out of what it judged. Rejected output never reaches the merge.
A run that only passes after retrying everything is not healthy — it is lucky. The console says which.
Wall clock against the same work done end to end. That ×factor is what the fan-out actually bought.
Walk a finished run's recorded outputs through the flow as it is drawn today — the canvas rings every block whose fate would flip. Decisions, not outputs; nothing is executed.
A run that stops for a human posts to Telegram with ✅ / ❌ buttons, and asks its questions as messages you answer by replying. The app polls the bot, so the decision and the answer come home with no public URL, no inbound port and no app to install — Slack works the same way through reactions and threads.
Wire the crew for the job.
The same sixteen blocks, arranged for the work in front of you. None of these needs a line of glue code — only a different drawing.
Ship from a ticket
Pull a story from Linear, split the backend and frontend work across sub-agents, and push the changes.
Audit a diff
A reviewer agent sweeps changed files for correctness bugs and security issues, ranked by severity.
Data-aware answers
An agent queries your database and answers with the real numbers — not a guess from training data.
Flows run while the app is running — not at 3am on a server.
That trade is the whole point. Driving the claude CLI on your machine is what reaches your Claude login, your repositories and the folders on your own disk — none of which a hosted container can see. A scheduled or mail-triggered flow fires when your machine is on and Concentus is open.
Install it. Open it. That is the setup.
The app carries its own Java runtime and its own PostgreSQL, so there are no prerequisites and nothing to configure — and it touches neither, if you already have them installed. The first launch asks you to create an account, and that is the whole of the setup.
Installer
A per-user install — no administrator rights needed.
SmartScreen showing a warning is normal: the Windows installer is not
code-signed yet. Click More info → Run anyway — or install through scoop,
winget or npx concentus, which skip the warning entirely.
Disk image
Signed and notarized — opens without the Gatekeeper dance. Drag Concentus into Applications.
AppImage or .deb
The AppImage runs as-is; the .deb installs like any other package.
chmod +x Concentus-*.AppImage
./Concentus-*.AppImage
# or
sudo apt install ./concentus_*_amd64.deb
Sign in to Claude once
To run flows on your subscription, sign in to Claude Code once in a terminal. Concentus finds the CLI itself.
claude # then /loginPrefer the cloud API? Set ANTHROPIC_API_KEY instead.
The interface speaks English, Spanish and Catalan — it follows your system language on first launch.
Build it yourself
Open source. Build the installer for your own platform, or run the app straight from the repo.
pnpm install
pnpm desktop # build + launch
# or produce an installer:
cd apps/desktop && pnpm packagePackage managers
No SmartScreen wall, and updates arrive the way the rest of your tools do.
scoop bucket add concentus https://github.com/Gergilcan/concentus && scoop install concentus
winget install concentus
brew tap gergilcan/concentus && brew install --cask concentus
npx concentusSign in to Claude once, on your machine. Concentus runs your flows through the claude CLI that login already belongs to — which is why there is no API key to paste. Anthropic meters non-interactive Claude Code use in its own allowance, separate from your interactive sessions; the Usage page measures what this machine has consumed, and a flow can fall back to an API key or a local model when it runs out.
Everything else, including the webhook setup, is in the README and the documentation.
Free for one. By seat past that.
One person, one licensee, forever free — the app also runs with no license at all, at the same one-seat limit. Team (up to ten seats, bought here) and Enterprise unlock the shared database, more than one member, and SSO.
Free
A license with your name on it — the app works without one too, at the same one-seat limit. Enter your name and email and it lands in your inbox.
Up to ten seats, bought here
What a team needs to work together, without the email: pick seats and term, pay by card, and the license lands in your inbox minutes after checkout. 25 € per seat per month; 20 € when you pay the year at once.
- Google and Microsoft sign-in
- Members and roles, up to your seats
- The shared database
Per seat, monthly or annual
Everything Team has, at any size, plus scale and governance. An invoice instead of a card, or a contract; annual billing is 10% cheaper. No prices are printed here yet — write in and we'll work out seats and term with you.
- Custom identity providers (any OpenID Connect issuer)
- Automatic accounts for an email domain
- OpenTelemetry metrics export to your collector
- Organization policies
- Audit trail export
- Unlimited retention of runs and versions
- Several organizations on one deployment
- Unlimited service accounts
- Published endpoints without a rate limit
- Groups inside an organization
Fourteen days of the team features, no card
A trial license for up to ten seats, in your inbox in a minute: the shared database, members, SSO — everything a team or enterprise license unlocks. When it ends the installation drops back to one seat; nothing is deleted, and a license picks up exactly where the trial left off. One trial per address.
Build your first flow in minutes.
Install the app, sign in to Claude once, and run a multi-agent flow on your own subscription — no API key, no setup tax, and a usage page that tells you where you stand.