Skip to content

A Kanban board your coding agents work, and you approve.

Works with Claude Code, Codex, Cursor and Gemini CLI.

kanbo is a Kanban board for AI coding agents that lives in your project. Agents take cards over MCP or the CLI; only a person can approve one.

$npm install -g kanbo-cli
GitHub Docs
Backlog1
WST-007

Warn when frost is forecast

To Do3
WST-004

Export readings as CSV

WST-005

Add a /export.csv endpoint

WST-006

Add a Download button to the dashboard

In Progress1
WST-002

Add a dark theme to the dashboard

writing the toggle; the colours are done, the Settings page is next

In Review1
WST-003

Fix humidity rounding on the daily chart

humidity rounding fixed, please check

Waiting for you
Done1
WST-001

Read the BME280 sensor over I2C

agent's shell · CLAUDECODE=1

$ kanbo card wait-approval WST-003 --text "humidity rounding fixed, please check"

your shell

$

how it works

Agents take the next card over MCP or the CLI

Tell your agent “take the next card from kanbo.” It takes a card, says what it is doing, and stops when it needs you.

01

It takes the next card

The first card in To Do that nobody is working on.

$ kanbo run start WST-002 --agent claude

To Do → In Progress · running: claude

02

It hands the card to you

The work is done: it moves the card to review, asks for a person and stops.

$ kanbo card wait-approval WST-003 --text "humidity rounding fixed, please check"

In Progress → In Review · waiting for you

approved, then moved on by the agent

03

You approve

In your own terminal or on the board page. Then the agent moves it on.

$ kanbo approve WST-003 --comment "Checked on the chart, looks right."

Only a person can approve

agent's shell · CLAUDECODE=1

$ kanbo approve WST-003
Only a person can approve a card. This shell belongs to an agent (CLAUDECODE=1): ask a person to approve it on the board page (kanbo serve).
If you are a person in an editor terminal, run: KANBO_ACTOR_KIND=person kanbo approve WST-003

exit 4

On Postgres, the kanban_agent role is refused by the database too: SQLSTATE 42501.

your shell

Human approval, enforced: exit 4 in an agent’s shell, SQLSTATE 42501 on Postgres

$ kanbo approve WST-003 --comment "Checked on the chart, looks right."
Approved WST-003

A board in a file, or a shared Postgres

one project

.kanbo/board.db

A SQLite file next to your code. Every command opens it directly; no server to run.

a team

Postgres

One board for people and agents. Supabase works.

In a terminal, or in your browser

$ kanbo boardweather-station
Backlog (1)
WST-007 Warn when frost is forecast
To Do (3)
WST-004 Export readings as CSV
WST-005 Add a /export.csv endpoint
WST-006 Add a Download button to the dashboard
In Progress (1)
WST-002 Add a dark theme to the dashboard [running: claude]
writing the toggle; the colours are done, the Settings page is next
In Review (1)
WST-003 Fix humidity rounding on the daily chart [waiting for you]
humidity rounding fixed, please check
Done (1)
WST-001 Read the BME280 sensor over I2C

The same board in your terminal, or in your browser.

$ kanbo board
$ kanbo serve → http://127.0.0.1:4318

The board page is the one at the top of this page.

agents

Works with Claude Code, Codex, Cursor and Gemini CLI

And anything else that can run a shell command or speak MCP over stdio.

16

kanbo_* MCP tools, and a kanbo command for every one.

connect one once

kanbo connect claude --yes

Writes the short section in CLAUDE.md and registers the MCP server in .mcp.json.

Measured, not promised

1,776tokens

kanbo_card_list at 1,000 cards, down from 178,117.

4.57ms

kanbo_ready on a 10,000-card SQLite board, down from 52.3 ms.

Apple M1 Pro, Node 24.13.0, p50.

264of 264

rule checks kept by Claude Code and Codex over 69 real runs (except one “at most two calls” rule in one scenario).

$0.21

a typical Claude Code “take the next card” task, median: 9 kanbo calls, 37 s.

Performance →

faq

Questions

Can an agent approve its own work?

Not through kanbo: there is no MCP tool for it, and the CLI refuses in a shell with KANBO_ACTOR_KIND=agent (exit 4). On Postgres with the kanban_agent role, the database refuses too. On a board file, an agent that writes to the SQLite file with other tools is not stopped — that is what the Postgres roles are for.

Do I need a server?

No. Every command opens the board directly: a SQLite file in the project, or a Postgres database. kanbo serve is optional, for the board page and the HTTP API.

Where is my data?

In .kanbo/board.db in the project (after kanbo init --file), or in the Postgres database you bound the project to. kanbo sends nothing anywhere else.

Does kanbo launch agents?

No. You or your orchestrator start agents; kanbo records their runs (kanbo run start / finish) so the board shows who is working on what.

Which agents work with kanbo?

Anything that can run a shell command or talk MCP over stdio: Claude Code, Codex, Cursor, and your own scripts.

MCP or CLI?

Either; they call the same operations. MCP gives the agent typed tools and is always treated as an agent. The CLI works for agents without MCP and for you.

Is there a hosted version?

No.

Does kanbo talk to GitHub?

No. kanbo card pr add records a link. The rule ci_green and the pull request standing on the board read facts written onto the card (as comments with dedupe keys) by whatever watches your pull requests; kanbo itself does not poll GitHub.

All questions →

your shell · install

Install in one command

$npm install -g kanbo-cli

$ cd your-project && kanbo

Then tell your agent: “take the next card from kanbo.”

Node ^22.19.0 || >=24.11.0 · macOS, Linux, Windows · free for noncommercial use, source on GitHub