Mission control for teams of
AI coding agents.
See every agent at a glance, keep them on-process with built-in guardrails, and get a git-native audit trail across Claude Code, Codex and Cursor.
No account, no database: docker compose up -d --build brings up the deck on your machine.
The real deck interface, running an example scenario. Click an agent, open a ticket, switch views. Nothing here talks to a server.
One deck across every coding agent
Run 3–5 agents in parallel and you lose the thread.
Spawning agents is the easy part: it's free and commoditized. Keeping track of them is where it falls apart.
Who's stuck?
An agent hit a wall ten minutes ago and is silently burning tokens, and you have no idea which one.
Who's waiting on you?
Three agents need a decision to continue. The answers are buried in terminals across four worktrees.
Who shipped what?
Something merged to main. No audit trail, no review record, no idea what guardrail did or didn't run.
A ticket-driven loop, recorded in git.
Every agent runs the same governed loop. State lives as YAML in the repo; the audit trail is git log.
Not another way to run agents in parallel, but a way to keep them honest.
The parallel-agents race is commoditized and free. We win on the layer above it.
Observability
One live dashboard across every agent, tool and project: states, tickets, who's stuck, who shipped.
Safety & human-in-the-loop
Approvals, awaiting-input alerts, and no --dangerously-skip-permissions by default.
Process governance
Validators enforce skills + evidence, status read-confirms and message banners. Almost no one does this.
Git-native audit · zero-DB
State is YAML in the repo, audit is git log. SOC2 / ISO-42001-friendly evidence for free.
Cross-tool, cross-project
Claude Code + Codex + Cursor + Kimi, many repos, one deck.
Where we actually win.
The category is crowded and mostly free. This is the table we're happy to be measured on.
Conductor · Vibe Kanban · Claude Code Agent Teams, as of May 2026. Corrections welcome via PR.
Built for the EU AI Act era.
Your agents shouldn't run with --yolo. agents-deck gives every action a reviewer, a guardrail and a git-native record: the evidence trail regulated teams need, generated for free.
Two commands and a browser tab.
The project itself is created in the wizard, not from a terminal. Everything stays on your machine, in your repository.
Before you start: Docker and git, no build step. The repository goes public soon; until then the clone URL comes with your early access.
Never run npx agents-deck. That npm name belongs to an unrelated third-party package, also published as ccdeck and agent-dag. This framework is only ever invoked by explicit path: node dist-server/cli.js.
Start the dashboard
one commanddocker compose up -d --build brings up the deck at localhost:4443. No database to provision, no account to create.
Add your project in the wizard
in the browserProjects → Add project, paste the absolute path to your repo. The wizard scaffolds project.yml, one folder per agent, their status files and system prompts.
Transmit the first ticket
one commandDraft a ticket, then hand it to an agent with the CLI: it moves from drafts to planned and the agent receives a message pointing at it.
The core is free, forever.
Individuals never pay. Teams who get outsized value fund development through governance.
- Full dashboard, tickets, messages, changelog, roadmap
- CLI + MCP
- Guardrails & human-in-the-loop
- Single user · self-hosted
- Community support
- Everything in Free
- Multi-user & roles
- SSO (Google / GitHub)
- Multi-project org dashboard
- Policy & guardrail packs
- Hosted option · email support
- Everything in Team
- SAML / SCIM · SOC 2
- Audit & compliance export (EU AI Act, ISO 42001)
- On-prem / air-gapped
- SLA & dedicated support
- Custom policy packs
agents-deck is free, and will stay free. Use it 100%, without paying a cent. If it earns its keep, back the project: pay what it's worth to you.
Questions, answered.
I'm already using Claude Code Agent Teams. Why agents-deck? +
Agent Teams lives inside a single Claude session and drives Claude only. agents-deck keeps the state on disk as YAML in your repository, shows every agent on one deck, and works the same way with Claude Code, Codex and Cursor. That reading matches the comparison table above, as of May 2026.
Is this just a prettier UI for Claude Code? +
No. It is a control plane: tickets that move through folders, guardrails that run before a commit, human approval gates, and a git-native audit trail. The dashboard is the visible part of it, not the product.
How is it different from Conductor or Vibe Kanban? +
Both are good at launching agents. What they do not carry is process governance and a git-native audit trail, which is the whole point here. The capability-by-capability reading is in the table above, dated May 2026, and corrections are welcome via PR.
Why does a git-native audit trail matter? +
Every ticket, message and status change is a file in your repository, so git log is the history: readable, diffable, reviewable in the tools you already use. No proprietary log format, nothing to export, and it outlives the session that produced it.
What does pricing apply to: per user, per agent or per repo? +
Per supervising developer, meaning the human piloting the agents. Not per agent, not per repository. Free is single-user and self-hosted, Team is 19€ per seat per month, Enterprise is custom.
Can I self-host it? +
That is the default, and today it is the only way to run it: you clone the repository, build it once and start the server yourself. There is no database, the state is YAML inside your repository, and nothing leaves your infrastructure.
Comparisons with other tools, as of May 2026. Corrections welcome via PR.
Get early access, and early-adopter pricing.
The first ~20 teams get −50% for life. Tell us what you need; we'll reach out before launch.
Check your inbox.
We've sent a confirmation link. Click it to confirm your spot.