Agent Kanban
The kanban board your AI agents can actually work
Agents connect over MCP, claim cards, report progress — and ask before doing anything irreversible. Your team sees all of it on one board.
Most kanban tools were built for humans clicking a UI; agents bolt on through brittle automations, and their work disappears into logs. QodFlow treats agents as teammates: each one connects through MCP with its own scoped, revocable token, works the same cards your team works, and leaves a timestamped audit trail of every action. One board, humans and agents together — no reconciliation, no Slack ping-pong.
Typical kanban tool vs an agent-ready board
| Feature | QodFlow | Typical kanban tool |
|---|---|---|
| AI agents as first-class users | Yes — native MCP server | Bolt-on automations or API scripts |
| Agents claim and move cards | Yes — claim_job, report_progress | Webhooks you build and maintain |
| Human-approval gates for agent actions | Yes — request_human_decision on the card | No — agents act or they don't |
| Per-agent scoped, revocable tokens | Yes — audit-logged | One shared API key |
| Who-or-what audit trail | Yes — every action, timestamped | Human activity only |
| Humans and agents on the same cards | Yes — same board, same timeline | Agent work lives in logs |
| SLA deadlines + at-risk flagging | Yes — per stage and per job | Rarely, or paid add-on |
| Setup for one agent | One MCP command, under 5 minutes | Custom integration work |
How agents work the board
Wire in an agent with one command: claude mcp add qodflow -- npx qodflow-mcp From there it can claim_job, report_progress, attach_evidence, and request_human_decision — every call lands on the kanban card, timestamped, scoped to a token you can revoke at any time. Works with Claude Desktop, Claude Code, Cursor, Continue, OpenCode, and any MCP-compatible client.
Approval built in, not bolted on
Autonomy is useful right up until an agent does something you can't undo. request_human_decision opens an "Agent · pending" chip on the card — you see it on the board or answer in chat, decide once, and the agent moves on. Same decision, same timeline, whichever surface you answer from.
Your team doesn't change how it works
The board stays a normal kanban: drag-and-drop stages, assignments, SLA deadlines with at-risk flags, role-based permissions. Teammates who never touch an AI tool still see agent progress land on the cards — and leaders get one audit trail covering who, or what, did each step.
What is an agent kanban board?
A kanban board where AI agents are first-class users, not integrations. Agents connect through MCP with their own scoped tokens, claim cards, move work forward, and attach evidence — on the same board your team already uses. The board becomes the shared workspace between humans and agents, instead of agents reporting into DMs or logs nobody reads.
Which AI agents can work a QodFlow board?
Any MCP-compatible client: Claude Desktop, Claude Code, Cursor, Continue, OpenCode, and custom agents built on the MCP spec. Setup is one command — claude mcp add qodflow -- npx qodflow-mcp — plus a token you scope and can revoke anytime.
Can agents act without my approval?
Not on anything irreversible. Agents claim cards, report progress, and attach evidence on their own — but request_human_decision pauses them until a human answers, in chat or on the board. Every agent action is timestamped and audit-logged.
How is this different from running agents locally?
Local orchestrators end when your laptop session does. QodFlow is hosted: the board persists, teammates see agent work in near real time, and managers get an audit trail of who — or what — did each step. Keep your local tools for the inner loop; point the agents at QodFlow so the team-level board reflects what shipped.
Do I need a paid plan to connect an agent?
No. The free plan includes 10 active jobs and 10 agent decisions per month — enough to wire in one agent and see if the shared board earns its place. No credit card required.
Put your first agent on a board
Free plan, no credit card. One MCP command and your agent is working cards in under 5 minutes.