Open source · Apache-2.0 · runs locally

You plan the work. Your agent reports back.

Vibeflow is agent task management made for humans: you plan on a kanban board in your repo — sequenced, claimed, reviewed. Your agent claims work, posts progress, flags what needs attention, and asks for human verification before anything lands.

No account No cloud Node 22+

kanban · acme-app Demo

Example data — drag cards, open tasks. Saved in your browser only.

Open full demo
  • Apache-2.0Open source, forever free in local mode
  • Local-firstTasks live in .vibeflow/, inside your repo
  • On npmnpx @vibeflow-tools/cli — nothing to install
  • You decideYour agent reports back — you verify before anything lands

Report a bug

Point at the bug. The task arrives actionable.

UI bugs used to arrive as prose — “the button looks off”. Annotate the element instead and the report carries everything your agent needs to act: what it is, where it lives, what it looked like. It lands on the board as a task that is ready to claim.

  • You annotate

    Press Alt+A, click the broken element, type one line. No extension, no account — one script tag or bookmarklet arms any page.

  • Context captured

    Selector, component, file and line number, screenshot and console errors — captured the moment you click, so the agent never re-derives them.

  • Ready to claim

    The report lands in todo with its context attached. Your agent claims it with tasks --next and posts progress back — you verify.

overlay → board
It lands in todo with its context attached — selector, component, file and line, screenshot. Your agent claims it with tasks --next, posts progress, and asks for human verification. That loop runs on the board above.

Task management

Your plan — and your agent's status channel.

You plan the work on a board you will actually open — with states, history, and child tasks that break big work into pieces. The same board is your agent's status channel: it claims work through the CLI or MCP, posts progress, flags what needs attention, and asks for human verification. UI bug? Point at it and the task arrives actionable — selector, file and screenshot attached.

  • Kanban board

    Backlog → todo → in-progress → review → done, with filters by status, tag, type and user, plus an activity feed per task. You plan here; your agent reports here.

  • Parent / child tasks

    Break an epic into children that keep their own status and history. The tree renders in the board and in the CLI — you break it down, the agent picks up a leaf.

  • Screenshots that attach themselves

    Paste a screenshot anywhere in the panel, or let Vibeflow capture the annotated element — it lands on the task, not in a Downloads folder.

  • Local-first storage

    Every task is a JSON file in .vibeflow/. Commit it, diff it, grep it, delete it. No database, no proprietary format, no lock-in.

  • CLI and board parity

    Every action exists in the terminal and on the board. You plan on the board; the agent claims from the CLI; review is the handshake — the agent asks, you verify. The same twelve tools are also served over MCP, over stdio or local HTTP — no account.

  • Safe for parallel agents

    --next claims atomically, so a swarm of agents can drain a backlog without two of them ever touching the same task. You watch progress land; nothing is done until you verify.

localhost:3700/kanban
The real board, served from the CLI on localhost — you plan here, your agent claims with tasks --next and reports back for verification.

How it works

Three steps. About ten seconds.

No extension, no account, no config file. One command starts a local server with the kanban board — and the annotation layer for the app you already have running.

  1. 01 You start

    One command. Board up, overlay armed.

    Start the local server and the kanban board opens — the same server serves the annotation overlay. Drop one script tag into the app you already have running, or drag the bookmarklet.

    Nothing is proxied and nothing is rewritten — the overlay is a single script that talks to your machine.

    terminal
    $ npx @vibeflow-tools/cli kanban
     
    <!-- then annotate your running app -->
    <script src="http://localhost:3700/vibeflow-overlay.js"
            data-vibeflow-overlay></script>
  2. 02 You point

    Click it. Name it.

    Hover to see the selector, click to pin a task, type one line, set a priority. Vibeflow captures the CSS selector, the React component and the source file and line number for you. Alt+A toggles annotation mode.

    Screenshots attach to the element they describe, and everything lands in .vibeflow/ where it can be committed to git.

    what lands in the task
    selector #signup-form > button.submit component SignupForm file src/auth/SignupForm.tsx:42 priority High
  3. 03 Agent reports

    Your agent claims it — and reports back.

    Paste one line into Copilot, Claude Code, Cursor or opencode. The agent claims the highest-priority todo task, moves it to in-progress, reads the full context and implements it.

    The claim is atomic, so several agents can run at once without ever picking up the same work. Progress, attention flags and verification asks land on the board where you watch — nothing is done until you verify it.

    paste into your agent
    Get next tasks and implement them:
    npx @vibeflow-tools/cli tasks --next
     
    ▸ claimed a3f2e1b8 · in-progress
    ▸ implement, commit, mark review

The problem

Your agent is fast. Managing its work is the slow part.

Agent-written code piles up faster than anyone can review it. Work lives in chat transcripts, context lives in the browser, and the task lives nowhere. Vibeflow puts every piece of work on one board — claimed, sequenced, and verified before it lands.

  1. 01

    Agent output with no status channel

    Your agent ships code faster than chat transcripts can track. Without a board that carries progress, attention flags and verification asks, nobody knows what is done, what is draft, and what needs a human.

  2. 02

    No home for the work

    Agent output lives in chat transcripts. The context lives in the browser. The task itself lives nowhere. Every follow-up starts from zero.

  3. 03

    Context gets re-derived

    Every hand-off re-derives the same context: which element, which file, which line. Ten minutes of detective work around a two-minute fix.

Bring your own agent

Works with the agent you already pay for.

No API keys, no plugin, no vendor lock-in. Tasks are plain files and the CLI is a single command, so anything that can read a file or run a shell can use Vibeflow. What comes back: the agent moves its own cards and asks for human verification when a task needs eyes.

  • GitHub Copilot
  • Claude Code
  • Cursor
  • Windsurf
  • opencode
  • Any CLI agent
  • MCP · 12 tools

The only prompt you need

Paste it once. --next claims the top todo task atomically, prints the selector, file and line, and moves it to in-progress — safe with several agents running at the same time. When the agent needs you, the card moves to review and asks for verification.

agent prompt
"Get next tasks and implement them:
 npx @vibeflow-tools/cli tasks --next"

Install

One command. You are building.

No sign-up, no dashboard, no onboarding. Run it once with npx, or install it globally for day two.

no installation required npx @vibeflow-tools/cli kanban

Annotate any page

localhost:3700/inject

Drag the annotate bookmarklet to your bar — no code change, any page.

Or serve a folder

npx @vibeflow-tools/cli serve ./my-app

The same overlay, injected for you — an alternative to the script tag or bookmarklet.

Install for day two

npm install -g @vibeflow-tools/cli

Adds the short alias: vf kanban

Requires Node.js 22+. Any framework — context is richest with React in dev mode. Licensed Apache-2.0.

Who it is for

Built for one developer. Ready for a team.

Solo developer

Free

You stay the decision-maker while your agent does the legwork.

  • Annotate your own prototypes and product pages
  • Drive your agent straight from the terminal
  • Tasks in git — no external service, no account
  • Works offline, and keeps working if Vibeflow disappears
  • Parallel agents drain the backlog for you
Free Local mode · Apache-2.0 · no account, ever

Team

Beta

Designers, PMs and devs working from the same board.

  • Non-developers annotate; developers ship
  • Shared board, comments, attachments, activity
  • Same tasks your agents read from the CLI
  • Plain files, so there is nothing to migrate back out of
  • Free for early users while it is in beta
Free while in beta Team cloud · pricing announced when it leaves beta

Questions

Frequently asked.

Do I need a browser extension?

No. Vibeflow injects its annotation layer with a script tag, its own file server, or a bookmarklet. No extension, no store approval, no browser permissions — and it works on any page, including a production build.

Do I need an account or a cloud service?

No. Local mode runs entirely on your machine and stores tasks as plain JSON files inside your project. Nothing is uploaded, there is no dashboard to sign in to, and the CLI works offline.

Will it work with my framework?

Yes. React, Vue, Svelte, Next.js, plain HTML — anything that renders in a browser. Context is richest with React in development mode, or any build that ships source maps, because Vibeflow can then resolve the component, file and line number for the element you annotate.

Where are my tasks stored?

In .vibeflow/tasks/ inside your project, as versioned JSON files. No database, no proprietary format, no lock-in — commit them to git, read them with any tool, or delete the folder and nothing breaks.

How is this different from a regular issue tracker?

Regular trackers plan human work. Vibeflow is the shared plan — and the status channel — between you and your agent: you plan on the board, your agent claims work, posts progress, flags what needs attention, and asks for human verification before anything lands. It stores plain JSON, so it can live alongside the tracker you already have.

Can several agents work in parallel?

Yes. tasks --next claims the highest-priority todo task atomically and moves it to in-progress, so two agents running at once never pick up the same work.

Is it really free?

Local mode is free and open source under Apache-2.0, and it always will be. A shared team version is in beta and free for early users; pricing has not been announced.

How does my agent connect?

Two ways, both local-first with no account. Over the command line, your agent runs npx @vibeflow-tools/cli tasks --next to claim work atomically. Or connect any MCP client to the built-in server: the same twelve tools over stdio or local HTTP. See the MCP setup guide.

Stop losing track of agent work.
Start managing it.

One command, no account, and the tasks stay in your repository.

Apache-2.0 · Node.js 22+ · no account required