Best Kanban Board for MCP in 2026: Seven Boards by Use Case

There is no single best kanban board for MCP. It depends on which AI client you use and what you need the agent to do. Trello is the cheapest paid option with OAuth and a ChatGPT route. Linear is the natural pick for engineering backlogs. Hypertask suits keyboard-heavy teams that also want a CLI and built-in AI. Weft is for people whose agents in ChatGPT, Claude and Cursor need to take tasks without colliding, then hand the result back for review on a phone.

We make Weft. Facts about other products come from their official pages, checked on 27 October 2026.

Hypertask's two posts on this topic put Hypertask first and say they rely on documentation, not hands-on tests. That is a fair method. This article uses the same method, but it ranks by use case and fills in the columns that matter once an agent is actually connected: how sign-in works, what happens in ChatGPT, whether an agent can claim a task over MCP, and what one seat costs.

What should a kanban MCP server do for you?

At minimum, an AI client should be able to find your board, read tasks, create one and move it between columns. Almost every tool below does that. The differences show up in five places:

• Sign-in. OAuth in the browser, or an API key you copy into a config file. Keys get pasted into chats and repositories, and an expiring key quietly breaks the connection.
• Your client. “Works with Claude” does not mean it works in ChatGPT. Some vendors have a ChatGPT app; others need a custom GPT or developer mode.
• Claiming work. If two agents can see the same task, something has to stop both of them from doing it. That can be inside MCP or only in a separate API.
• Where you check results. A web board, or also a phone app that tells you when an agent needs an answer.
• Price per seat. Plus any AI credits the board itself sells.

How do the seven boards compare?

Prices are USD per seat per month on the cheapest paid plan. Token lifetimes are listed where vendors state them.

Two things stand out. Every board here has an official hosted server, so “has MCP” no longer separates anyone. And in what we checked, only Weft exposes claiming as an MCP tool; Hypertask has leases but keeps them in its REST API. We keep a longer, regularly updated table of MCP task managers.

Which board should you pick?

When is Trello, Linear or Hypertask the better choice?

Trello is the better choice if price and familiarity matter most. A seat costs $5 a month billed yearly, sign-in is OAuth only, MCP works on the Free plan, and it is reachable from ChatGPT and listed in Claude's directory. Weft is not in Claude's directory yet; in claude.ai you add it as a custom connector. Trello's server also now connects several Workspaces at once.

Linear is the better choice for software teams whose source of truth is already a list of issues, cycles and projects. Weft has sprints and projects, but no GitHub integration, so pull requests do not update cards on their own. See Weft vs Linear.

Hypertask is the better choice if you want a CLI, a REST API or AI built into the board. Weft has none of those three: no CLI, no public API, no webhooks, and no AI of its own. In Weft the thinking happens in your AI client, which is why there are no AI credits to buy. Hypertask's MCP has 21 tools (tools reference); Weft's has 32.

Why does claiming a task through MCP matter?

Picture two chats, one in ChatGPT and one in Claude Code, both told to “pick up the next task.” If the board only offers a move-to-Doing call, both can read the same card as free and both start. You find out when two pull requests arrive.

In Weft, start_task claims the task with a 20-minute lease and moves it to Doing. The database decides who wins, so the second client gets already_claimed and moves on. report_progress renews the lease while the agent works. If the agent dies and goes quiet, the lease lapses and the task is free for someone else, instead of sitting in Doing forever.

When the agent hits a question only you can answer, request_input puts it in your Inbox, with options, and can send a push to your phone while the agent keeps its claim. complete_task records the outcome, links as evidence, what was not done and what is still unknown. By default a finished task goes straight to Done. When review is needed, or a verification check fails, it waits in your Inbox for Accept, Rework, Later or Drop.

There is also a guard on your attention. Once 5 items wait on one person (10 in a team workspace, or 45 minutes of review), start_task refuses new work with attention_queue_full until you catch up. The full loop is described in how the claim loop works, and running several clients at once in multiple AI agents on one board.

What does Weft not do?

• No CLI, no public REST API and no webhooks.
• No GitHub, GitLab or Slack integration.
• No AI built into the board and no AI credits; you bring your own client.
• Not listed in Claude's connector directory yet; claude.ai and Desktop use a custom connector.
• No desktop app; web, iOS, Android and a VS Code extension.
• Restoring a task from the trash is a Pro feature. Agents can trash tasks but cannot delete them for good.

How do you test any kanban MCP in 10 minutes?

Do this on a throwaway board before you trust any tool, ours included.

1. Read-only first. Connect the client you actually use and note how sign-in works: a browser consent screen, or a key you paste. If the board offers a read-only connection, pick it (Weft lets you choose full, write or read on the consent screen; Linear has a separate read-only endpoint). Ask: “List my boards and the tasks in the first column. Change nothing.” Check the answer against the board.
2. One reversible write. Switch to write access and ask for one change you can undo by hand: create a card called “MCP test” and move it one column to the right. Then ask the agent to move it back. If you run two clients, ask both to start the same card and see what the second one is told.
3. Check the board on your phone or the web. Open the board outside the chat. Is the card where the agent said it is? Does it show which client changed it? Then disconnect the client in the board's settings and confirm the next call fails. In Weft, Disconnect refuses the client's next call.

Setup pages for each client are at Claude Code, ChatGPT, Cursor and Codex. Moving existing work across is covered on migrate from another tool: Weft imports Linear, Jira and Notion CSV exports and Trello JSON on every plan.

Isn't Trello cheaper and good enough?

Often, yes. At $5 a seat a month billed yearly, with OAuth and MCP on every plan, Trello is a good default for a simple visual board (Trello pricing). Weft costs the same per seat on yearly billing ($60 a year) and adds claiming over MCP and a phone Inbox for agent questions. If you do not need those, stay on Trello.

Do I need to copy an API key to connect Weft?

No. Weft uses OAuth in the browser and issues refresh tokens, so the client renews sign-in itself. Personal access tokens exist for clients that cannot do OAuth; they are optional and you choose their expiry.

Does Weft work in ChatGPT on the Free plan?

Yes. Weft is a published app in ChatGPT's directory on every plan, including Free. Enterprise and Edu need an admin to enable it. See Weft for ChatGPT.

Is Weft in Claude's connector directory?

Not today. In claude.ai and Claude Desktop you add it as a custom connector with the URL https://letsweft.com/api/mcp. In Claude Code it is one command: claude mcp add --transport http weft https://letsweft.com/api/mcp.

What does the Weft Free plan include?

50 active tasks and 200 AI tool calls a month. Pro is $7 a seat monthly or $60 a seat yearly, with unlimited tasks and calls (pricing). There are no AI credits, because the AI runs in your own client.

I need a CLI for CI jobs. Is Weft right?

No. Weft has no CLI and no public API. Hypertask has an npm CLI and a REST API (CLI reference), and Linear is strong for engineering workflows. Weft fits when agents work from AI clients over MCP.