planban
Planban is a local-first, agent-native Kanban board that installs as a plugin in Codex and Claude Code, so your agents keep the plan current and you see the whole project at a glance.
It gives you a simple roadmap board, keeps your planning state local, and helps your coding agent start work from the right card, spec, and project context. Its MCP server also works from any host that can run a local MCP server.
Planban is free to use on your computer with no account. When you want to share a board, use it from another computer, or reach it from ChatGPT or claude.ai, you can take that one board online with a free Planban account (see Online Boards). Every other board stays on your computer.
Planban's broader product direction is a durable planning system shared by humans and agents, designed first to give an individual clear sight of project work across agent surfaces. Read the Planban Product Constitution for the principles that guide product and feature decisions. Learn more at https://planban.ai/.
What Planban Does
Opens local project boards beside your session with
/pbor/planbanin Codex, and/planban:pbor/planban:planbanin Claude Code.Keeps live planning state on your machine, separate from your source repo.
Gives agents stable board, card, spec, plan, and next-action context.
Organizes related outcomes into root Groups while each owned Item remains independently completable, ranked, documented, and movable.
Supports board creation, duplication, archiving, restoring, card movement, docs, history, and feedback flows.
Ships CLI, local API, and MCP tools so any agent host can work with Planban through structured local actions.
Takes a board online when you choose, so people you invite and their agents can work on it, with roles and limits on what agents may do.
Related MCP server: productplan-mcp-server
Install
Planban is installed separately in each host you use. Each install shares the same device-local boards, so a board you create from Codex is the same board you open from Claude Code.
Host | Slash commands | Board beside the session | Install |
Codex desktop |
| Codex in-app browser | |
Claude Code desktop |
| Claude Code browser pane | |
Other MCP hosts | none | link only |
When a project uses Planban but Planban is not installed in the host you are working from, the project's .planban/agent-context.md tells the agent to say so once and point you here.
Install With Codex
Ask Codex:
Install Planban from planban-ai/planban. Follow the Install With Codex details in the public GitHub README exactly, verify the plugin and MCP tools work, open the interactive tutorial in the Codex in-app browser, then ask whether I want to set up Planban for a local project.Install With Codex details
These details are written for Codex to follow. If you want to run the commands yourself, use Manual Install below.
Planban installs as a Git-backed Codex plugin marketplace. Codex should:
Check whether Node.js and npm are available. Planban requires Node.js
>=22.12.0so its local runtime and build tooling stay inside supported dependency ranges. If Node.js or npm is missing or too old, explain that Planban runs locally and requires a modern Node.js LTS release. Ask before installing or upgrading Node.js. If approved, install Node.js LTS using the safest method for the machine, then verifynode --versionandnpm --versionbefore continuing.Add the Planban marketplace.
Locate the installed marketplace root.
Run
npm installthere.For a local-clone install, configure the Planban MCP runtime before installing the plugin. Git-backed marketplace installs can use the shipped relative MCP bootstrap without configuration.
Install the Planban plugin.
Verify the plugin and MCP tools are available.
Open the interactive Planban tutorial in the Codex in-app browser.
Ask whether to set up Planban for one of the user's local projects.
Manual Install
Use this path only if you prefer to install without asking Codex to perform the setup.
Recommended Git-backed marketplace install:
codex plugin marketplace add planban-ai/planban
PLANBAN_ROOT="$(codex plugin marketplace list | awk '$1 == "planban" { print $2 }')"
cd "$PLANBAN_ROOT"
npm install
node scripts/configure-local-plugin.mjs "$PWD"
codex plugin add planban@planban
codex plugin list --marketplace planban
node plugins/planban/scripts/launch-planban.mjs --tutorialThen open the printed local tutorial URL. In Codex, ask your agent to open it in the in-app browser.
For a Git-backed marketplace, the shipped plugin can also bootstrap itself: after
adding the marketplace and installing planban@planban, start a new Codex session
and ask to open Planban. First launch installs missing dependencies in the
marketplace runtime. Node.js and npm must already be available to Codex. The MCP
startup allowance is two minutes; if a slow network exceeds that allowance, finish
npm install in the marketplace root and retry. Installer output goes to stderr,
keeping MCP's JSON-RPC channel clean.
PowerShell manual install:
codex plugin marketplace add planban-ai/planban
$planbanRuntime = (codex plugin marketplace list --json | ConvertFrom-Json).marketplaces | Where-Object { $_.name -eq 'planban' } | Select-Object -ExpandProperty root
Set-Location $planbanRuntime
npm install
node scripts/configure-local-plugin.mjs "$PWD"
codex plugin add planban@planban
node plugins/planban/scripts/launch-planban.mjs --tutorialIf an older cached plugin still contains __PLANBAN_REPO_ROOT__, refresh the
marketplace and reinstall the plugin, then start a new Codex session. Opening a
Board alone does not refresh an already-running session's MCP configuration.
On Windows, Planban uses polling for its small local planning state directory. This avoids native file-watcher permission failures while keeping board updates live. You can also force polling on other platforms with CHOKIDAR_USEPOLLING=1.
Local clone fallback:
git clone https://github.com/planban-ai/planban.git
cd planban
npm install
node scripts/configure-local-plugin.mjs "$PWD"
codex plugin marketplace add "$PWD"
codex plugin add planban@planban
node plugins/planban/scripts/launch-planban.mjs --tutorialInstall With Claude Code
Ask Claude Code:
Install Planban from planban-ai/planban. Follow the Install With Claude Code details in the public GitHub README exactly, verify the plugin and MCP tools work, open the interactive tutorial in the browser pane, then ask whether I want to set up Planban for a local project.Install With Claude Code details
Planban installs as a Claude Code plugin marketplace. Node.js >=22.12.0 and npm must be available. Then:
claude plugin marketplace add planban-ai/planban
PLANBAN_ROOT="$HOME/.claude/plugins/marketplaces/planban"
cd "$PLANBAN_ROOT"
npm install
claude plugin install planban@planban
claude mcp list
node plugins/planban/scripts/launch-planban.mjs --tutorialclaude mcp list should show plugin:planban:planban as connected. Start a new Claude Code session, then type /planban:pb in a project to open its board, or /planban:planban-tutorial for the tour. Claude Code opens the board in its browser pane when the desktop app provides one, and always replies with the clickable board URL.
Run npm install in the marketplace root before the first session. The plugin can install missing dependencies on first launch, but Claude Code allows MCP servers only a short startup window, so a slow network can make the first connection fail until dependencies are present.
Local clone fallback:
git clone https://github.com/planban-ai/planban.git
cd planban
npm install
claude plugin marketplace add "$PWD"
claude plugin install planban@planbanUpdates: claude plugin marketplace update planban then claude plugin update planban@planban, and run npm install in the marketplace root again when a release changes dependencies.
Claude Code copies the plugin's skills, hooks, and scripts into its plugin cache at install time, while the Planban runtime itself resolves to the marketplace checkout. claude plugin update refreshes that cache only when the plugin version changes, so when developing Planban from a local clone, run claude plugin uninstall planban@planban then claude plugin install planban@planban after changing anything under plugins/planban/, then start a new session; an idle session keeps its plugin server process alive and will not see the change until it restarts.
Other hosts
Any host that can run a local stdio MCP server can use Planban's tools without slash commands. Clone the repository, run npm install, and point the host's MCP configuration at node <checkout>/plugins/planban/scripts/start-planban-mcp.mjs with the environment variable PLANBAN_REPO_ROOT=<checkout>. The agent then gets the same tools and the same clickable board URLs; opening the board beside the session depends on what the host offers.
First Run
Planban opens a short local tutorial and creates a Planban Demo board so you can try the product immediately.
Use it to:
learn how
/pb,/planban, and/planban-tutorialwork in Codex, or/planban:pb,/planban:planban, and/planban:planban-tutorialin Claude Codedrag cards between columns
drag one Item onto another to create a Group, or move an Item into an existing Group
open a Group to review its objective, progress, and status-ranked Items together
open a roadmap item with your agent
inspect roadmap item details and specs
mark a card Complete when you are done
send feedback from the toolbar
ask your agent to create roadmap items from your existing plans
When you are ready to use Planban with a real project, ask your agent in Codex or Claude Code:
Set up Planban for my local project at /path/to/project. If it is not initialized yet, ask me before initializing it. Then open the board and help me create roadmap items from the current repo docs, issues, notes, or my description of what I am building.You can also give your agent planning context from Notion, Jira, Linear, GitHub Issues, copied notes, or a plain-language project update. Ask it to turn that context into draft Planban roadmap items for review.
Feedback With Your Agent
Planban uses GitHub Issues for bugs, feature requests, and product feedback.
Feedback is welcome. If you want to share a bug, request, rough edge, or reaction, select the feedback button in the board toolbar and add whatever you remember. Your agent will use that note plus available conversation and workspace context, gather relevant Planban diagnostics, search existing issues and fixes, and recommend the best reporting route before anything is filed publicly. Missing details do not block a report.
For bugs, Planban Feedback normally prepares a new issue or a useful comment on an existing issue. It recommends a pull request only when it can verify that you or your agent already implemented and validated a focused fix for the same failure. Reporting a bug does not by itself create or recommend a PR.
You can also start from a chat: use /planban-feedback in Codex or /planban:planban-feedback in Claude Code, or ask your agent:
I want to give feedback on Planban. Use the current conversation and safely discoverable Planban context so I do not have to repeat details. Investigate the report, search existing issues and fixes, then recommend and draft the best route for planban-ai/planban. Do not recommend a PR unless you verify that my agent already implemented and validated the fix. Do not include private repo paths, board contents, logs, screenshots, or personal project details unless I explicitly approve them.
Feedback:
<paste your feedback here>Your agent can then help you review the issue draft and file it through GitHub. You can also open the issue chooser directly:
https://github.com/planban-ai/planban/issues/new/choose
Contributing
Planban is early. Small focused bug fixes, docs improvements, and reliability fixes are welcome. For larger feature, architecture, install, update, API, MCP, or CLI contract changes, please open an issue first.
By submitting a pull request, you agree that your contribution is licensed under the MIT License.
See CONTRIBUTING.md, SUPPORT.md, and SECURITY.md.
Updates
Planban checks public version metadata from GitHub while the local board is open. If an update is available, the board shows a small update notice.
When Planban can prove your install is safe to update directly, choose Update now. Planban will show progress, refresh the local install, restart the local server, and reopen the board you were viewing.
If the install shape is ambiguous, dependencies are missing, a migration is needed, or local files make direct update unsafe, the board shows the Update with Codex panel instead. It prepares a prompt asking your agent to inspect the install, update Planban safely, verify the plugin and MCP tools, and open the right post-update surface. Open it as a Codex draft, or copy it into an agent session.
Fresh installs start with the interactive tutorial, which opens in the Codex in-app browser or the Claude Code browser pane. If you used an earlier private or preview build, use the agent-guided update or reinstall from this public repository so your local plugin marketplace points at the clean public release.
Manual updates
Codex, for the recommended Git-backed marketplace install:
codex plugin marketplace upgrade planban
PLANBAN_ROOT="$(codex plugin marketplace list | awk '$1 == "planban" { print $2 }')"
cd "$PLANBAN_ROOT"
npm install
node scripts/configure-local-plugin.mjs
codex plugin add planban@planbanFor a local clone install, run git pull in the clone instead of the marketplace upgrade, then the same remaining commands.
Claude Code:
claude plugin marketplace update planban
claude plugin update planban@planbanRun npm install in the marketplace root again when a release changes dependencies.
After a manual update, start a new session so the host loads the updated plugin and MCP server.
Planban does not silently update itself and does not send private board contents, repo paths, logs, or project details as part of update checks.
Online Boards
An Online Board lives at app.planban.ai instead of on one computer. Use one when you want to work on a board with someone else, open it from another computer, or let ChatGPT or claude.ai read and update it. Boards stay on your computer unless you take them online, and Planban never needs an account for local boards.
In the commands below, planban means Planban's command line. If it isn't installed as a command on your computer, Planban's own messages print the exact form to run, which is node "<Planban install folder>/bin/planban.mjs".
Take a board online. In the board, open Board settings and choose Take online, or run:
planban online login https://app.planban.ai # once per computer; opens sign-in in your browser
planban online take-online https://app.planban.ai --cwd /path/to/projectA board goes online only when it holds no files (see Local Storage). Taking it online leaves .planban/online-board.json in the project folder: commit it, so other computers and agents know the board lives online. This computer keeps a backup copy. You can bring the board back any time with Board settings, Take local (or planban online take-local), and its history comes with it.
Invite people. Board settings, People, Invite people, or planban online invite --email <address> --role editor (the role is viewer unless you say otherwise). Each person has a role:
Owner can do everything, including inviting people and taking the board online or local.
Editor can change cards, Specs and Plans, and mark work complete.
Viewer can read the board and its history.
Each person also chooses what their own agents may do: read, plan, or plan and complete (planban online my-agents <level>). An agent can never do more than its person. People join from the invitation link with planban online join <link>.
Connect AI apps. Codex and Claude Code with the Planban plugin already work with online boards once this computer is signed in. Other apps connect to https://app.planban.ai/mcp and ask you to sign in and click Allow:
ChatGPT and claude.ai: add a custom connector with that address.
Claude Code without the plugin:
claude mcp add --transport http planban https://app.planban.ai/mcp, then sign in with/mcp.Codex without the plugin:
codex mcp add planban --url https://app.planban.ai/mcp, thencodex mcp login planban.
See and cut off any of these from Board settings, Connections, or planban online connections https://app.planban.ai.
Limits. An online board holds up to 256 MB, each account can make up to 5,000 changes and write up to 64 MB of text a day, and Planban stays an index, not file storage.
Local Storage
Planban keeps live planning state on your machine:
repo discovery files:
.planban/project.jsonand.planban/agent-context.mdlocal board state:
~/.planban/repos/<repo-id>/roadmap.jsonlocal card docs:
~/.planban/repos/<repo-id>/items/<card-id>/spec.mdandplan.md
Planban is an index, not file storage. A card holds only text: its fields, Spec, Plan and history. Files stay where they belong, such as your repository or a shared drive, and a card refers to them with a GitHub link, a web address, or a local path labelled with its machine (Mac mini: ~/Projects/app/renders/front.png). Planban refuses text that embeds a file, Specs and Plans over 1 MB, and cards over 128 KB.
If you upgrade with files already in a card's folder, Planban flags that card and refuses changes to it until the files are moved out. Reading it still works. To tidy it:
planban move-files-out <card-id> --to <folder> --dry-run # preview
planban move-files-out <card-id> --to <folder>This moves the files to that folder, never overwriting or deleting anything, and adds a Files section to the card's Spec that refers to each one where it now lives.
Do not commit ~/.planban state to your project repo. We recommend adding .planban/agent-context.md to your project's .gitignore: planban init does not do this for you, and that file contains absolute paths from your machine. If the board is online, do commit .planban/online-board.json, so other computers know where the board lives.
License
Planban's local core is open source under the MIT License. See LICENSE.
The Planban name, logo, wordmark, domain, social handles, and official release
channels are not granted under the code license. See TRADEMARKS.md.
This server cannot be deployed
Maintenance
Related MCP Connectors
MCP server for generating rough-draft project plans from natural-language prompts.
MCP server for medicaid-intelligence
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceMCP server for toolhouse.ai. This does not rely on an external llm unlike the official server.4MIT
- AlicenseAqualityAmaintenanceMCP server for ProductPlan - enables AI assistants to interact with roadmaps, OKRs, and discovery features.503MIT
- AlicenseBqualityDmaintenanceMCP server for Maasy AI Marketing Copilot2121 npmMIT
- AlicenseCqualityDmaintenanceMCP server for Plane integration, enabling AI agents to interact with Plane APIs and services for project management tasks.100MIT