Implement features from ticket to PR automatically, in secure dev containers

Focus on defining your functional and non-functional requirements, then orchestrate each ticket's implementation while staying in control.

Get started Read the docs

Up and running in three commands

  1. Install the CLI (Homebrew on macOS / Linux)

    brew install takuto-team/tap/takuto

  2. Generate config with the interactive wizard

    takuto setup

  3. Bring it up, then open the dashboard (default :8080)

    takuto start

Prerequisites: Docker or Podman, a GitHub PAT, and an AI provider API key. takuto setup creates a .takuto/ folder anywhere; takuto start pulls the Takuto Core image and opens the dashboard.

Takuto flips the usual AI workflow: instead of prompting your way forward one step at a time, you write a proper spec up front, as detailed as you'd give a teammate. Takuto then implements it, either from tickets it pulls automatically or ones you add in the dashboard. This trades improvised vibe coding for deliberate planning, while you keep full control: drop into a terminal or IDE, or prompt the agent whenever you need. Every run stays sealed in its own container, so a prompt injection can't escape.

Demo edited: the ticket-description editing step was cut, and the video sped up.

Bring your own coding agent — Takuto drives these four:

  • Claude Code
  • Cursor Agent
  • Codex
  • OpenCode self-hosted models only (LM Studio, Ollama, vLLM…)

Where it came from →

Automation you stay in control of

Run the whole pipeline unattended, or take the wheel ticket by ticket. The choice is yours on every run.

Ticket-driven or standalone

Poll Jira or GitHub Issues for "To Do" tickets, or paste a description straight into the dashboard — no ticketing system required.

Autonomous, manual, or mixed

Let it run the full pipeline overnight, trigger each phase yourself, or auto-pick routine work while you curate the tricky tickets.

Parallel by design

Run multiple tickets at once — each gets its own git worktree and isolated environment, so nothing steps on anything else. Concurrency is yours to set.

A live dashboard

Stream terminal output per workflow, watch progress, and pause, resume, retry, or inspect any run from the browser.

Editor + terminal in the browser

Jump into any workflow with a VS Code editor and web terminal pointed at the exact worktree the agent is working on.

Pipelines you define

Chain steps — implement, address PR comments, merge the base branch — with dependencies, all edited in the dashboard’s Workflows tab.

Two ways to run it

Takuto CLI

The fastest path: takuto setup generates your config in a .takuto/ folder, then takuto start brings everything up and orchestrates Docker or Podman Compose for you — the dashboard is at :8080 by default. You finish setup (admin account, AI provider, GitHub) there.

Build your own from Takuto Core

Run the core engine directly — clone Takuto Core, configure config.toml, and bring it up with Docker Compose yourself. For teams that want full control of the image and deployment.

Isolation & privacy

Each agent runs in its own container

Work is isolated per workflow — separate container, separate git worktree, separate environment — which reduces the prompt-injection blast radius. Outbound traffic is locked down by a default-deny egress firewall. No tracking, no telemetry: your code and ticket content go only to the AI provider you configure.

The name

Takuto (タクト) — a conductor’s baton

In Japanese, takuto is the baton a conductor uses to lead an orchestra. The name fits what the tool does — keeping a section of AI agents in time, working from one score — and it’s a nod to Japan, where Takuto was built.