agentodo

A local-first kanban board for the work you hand to AI agents.

agentodo turns coding prompts into structured tasks on a real board. Assign an agent, pick a model, point it at a local project — then step away while it runs. Every run keeps its transcript, so you can review, follow up, or re-run from the same card.

Quick start

npx agentodo

That's it. The CLI boots the server at http://localhost:3737 and opens your browser.

npx agentodo --port 4000 --no-open

Prerequisites

• Node.js >=22.5.0
• At least one agent CLI installed locally (optional — needed only to run tasks):
  • claude (Claude Code)
  • codex (OpenAI Codex)

You can open the board and create tasks without an agent installed — you just can't run them.

Why it exists

Most AI coding workflows fall apart the same way:

• Useful prompts get lost inside long chat threads.
• Long-running work has no clear owner or status.
• Model, effort, and mode choices aren't tracked.
• Analysis and implementation get tangled together.
• There's no reliable record of what the agent actually did.

agentodo gives that work a shape:

• Create tasks instead of one-off prompts.
• Assign the right agent, model, and mode per task.
• Run against a selected local project directory.
• Move cards through Backlog → Todo → In Progress → Completed.
• Open the live console during the run, review the full transcript after.

Features at a glance

• Kanban board with a backlog panel for queued ideas.
• Two agents out of the box: Claude Code and Codex.
• Ask mode for read-only analysis · Code mode for implementation.
• Per-task controls: model, effort tier (Low → Max), Fast mode, task type.
• Live run console with streamed thoughts, tool calls, and shell output.
• Task-scoped transcripts — every run is attached to its card.
• Local project picker with a safe ~/.agentodo/scratch fallback.
• Keyboard shortcuts: N new task, B backlog, / search, Esc close.
• Light / dark themes with no-flash theme boot.

Your first task

1. Open the board at http://localhost:3737.
2. Click Add task (or press N).
3. Pick a project directory — or leave it empty to use the scratch workspace.
4. Choose the agent, model, effort, and mode (Ask vs Code).
5. Drag the card to In Progress (or hit run from the editor) to kick it off.
6. Watch the live console stream the agent's work.
7. Move the card to Completed when you're happy, or send a follow-up message on the same transcript.

Tip. Use Ask mode first to scope a change ("Trace how auth works in this repo"). Convert the result into a Code mode task once the direction is clear.

Local data

agentodo stores everything on your machine:

• SQLite database: ~/.agentodo/agentodo.db
• Scratch workspace (when no project is selected): ~/.agentodo/scratch

Nothing is sent to a remote server.

Development

git clone https://github.com/sushil-kamble/agent-todo.git
cd agent-todo
pnpm install
pnpm dev

• Client → http://localhost:3000
• Server → http://localhost:8787

pnpm build       # build shared + client for production
pnpm start       # run the built app via the root CLI (port 3737)
pnpm test        # run all package test suites
pnpm typecheck   # typecheck every package
pnpm check:ci    # typecheck + biome + tests

Tech stack

• Monorepo managed with pnpm
• Node.js >=22.5.0
• Client: TanStack Start, Vite, React 19, TypeScript, Tailwind v4
• Server: Node ESM HTTP backend with agent orchestration
• Shared: contracts, config, constants, runtime helpers, validation

License

MIT © Sushil Kamble