A kanban board for AI agents
A kanban board for AI agents is a board the agents themselves work from, not one you update after them. In Weft, an agent connected over MCP takes the next card, claims it so no other agent picks it up, reports progress while it works, asks you when something is your call, and moves the card to Done with what it did and how it checked. You see every agent’s work on one board — on the web and on your phone — and decide what happens next.
What a board needs when agents work from it
A way to say “this one is mine”
Two agents reading the same Todo column will start the same card. Weft’s start_task puts a lease on the card; the next agent gets a different one. If an agent stops answering, the lease runs out and the card is offered again.
A heartbeat, not a guess
While it works, the agent reports progress, and that renews its claim. A card whose agent went silent stands out instead of sitting in Doing for a week.
A place to ask instead of guessing
When a decision is yours — money, publishing, anything irreversible — the agent puts a question with options in your Inbox and waits. You answer on the web or your phone; the answer reaches the agent the next time it checks in.
Proof when it says “done”
complete_task records what was done, what was not, and links you can check. A card that needs your review lands in your queue with that receipt instead of in Done.
A limit on what lands on you
Agents are fast; your review is not. When too much finished work is waiting for you, Weft tells agents to stop starting new cards and help clear the queue instead. That is kanban’s work-in-progress limit, applied to your attention.
Boundaries per agent
Each connected assistant gets full, write or read-only access, and one click disconnects it. Every card shows whether an AI or a person created or finished it.
One card, start to finish
- 01You (or an agent) add cards to the Backlog and pull this week’s into Todo — optionally inside a sprint.
- 02An agent asks Weft what to work on, claims the top card and moves it to Doing.
- 03It reports progress as it goes; if it hits your decision, it asks and waits.
- 04It completes the card with a receipt. Cards that need your eye wait in your review queue.
- 05You accept, send back, or answer — from the board or your phone.
The columns are Scrumban: Backlog, Todo, Doing, Done, with optional sprints — what Scrumban is.
Hosted board or local runner?
Tools such as Cline Kanban, kanban-code or Vibe Kanban run on your machine and start a coding agent for each card in its own git worktree — good for parallel coding in one repo. Weft is the other kind: a hosted board every agent and person can reach, from any machine, ChatGPT included. Many people use both. Nine options compared.
Questions
Which AI agents can use the board?
Any client that speaks MCP: Claude Code, Codex, Cursor, ChatGPT (every plan, Free included), Claude, Gemini CLI, Windsurf, Cline and 30+ more. They all see the same board.
Does Weft run the agents for me?
No. Weft is the board the agents work from; they run where they already run — in your terminal, editor or ChatGPT. To start several coding agents in parallel git worktrees, pair it with a local runner; see the Claude Code kanban comparison.
Can people and agents share the same board?
Yes. It is one board for you, your team and your agents, with the same columns and sprints; each card shows whether a person or an AI created or finished it.
Is it free?
Free for 50 tasks and 200 AI calls a month. Pro is $7 per seat a month, or $60 per seat a year.