Early access

A colleague that works your queue.

Point kugutsu at a ticket queue, an inbox, a schedule or a webhook. It picks up what arrives, does the work with the models you choose, and hands you the result to approve, within the budget and the permissions you set.

Find what it could do for you Set it up with your agent

The kugutsu app showing a running graph: a webhook feeds three writer models in parallel, an editor model merges their drafts, and the result waits at a review step; on the right, the review with Pass and Reject buttons. — A launch team as a graph: three writers in parallel, an editor, then the copy waits for you to pass or reject it.

Work that starts without you typing

A kugutsu graph keeps running on your machine. Events reach the models, the models do the work, and what matters waits for you.

Something happens

Events

A webhook
A schedule
A file changes
A ticket moves
You type

The work

Agents and rules

Claude Code, Codex sessions
Claude, GPT, DeepSeek, local models
In parallel, in loops
Rules between them

You decide

Review

Pass or reject
From the app or a ticket reply
Everything waiting, in one inbox

The result

Effects

A pull request
A comment, a message
A file, a report
Only what you granted

Budgets stop it when spentEvery step and its cost recordedGraphs can reshape themselves, with your approval

Safe to leave running

It stops at your budget

Declared in the graph or for the whole machine, per day or per run. Once spent, no node starts another step.

It does only what you granted

A file written, a request made, a command run through kugutsu: each is granted to a node and checked before it runs. An agent session acts on its own, inside the boundary you set for it: files, network, commands, CPU and memory, sandboxed.

It shows you what it did

Watch messages move through the running graph, replay past runs, and test a changed graph against them before it goes live.

The kugutsu Inbox listing two reviews from two graphs, an urgent alert to escalate and launch copy to publish, each with Pass and Reject buttons. — One inbox for everything waiting on you, across graphs.

A recorded run in kugutsu: 14 steps in 37.6 seconds for $0.08, and the editor step's cost broken down by model tokens, bytes written and time. — Every run recorded, with what each step cost and why.

Which of your work would it run?

That is the hard question, and your own coding agent can answer it better than a list of examples. Two hand-offs, each one line to paste into Claude Code, Codex or another agent.

Step 1 · no install needed

Find your use cases

Your agent asks which sources it may read.
It looks for repetition in your agent and shell history, your commits and your existing scripts.
It asks you a few questions about your week.
It writes a report of candidates: what would start each, the steps, where you review, what a mistake would cost.

Say to your agent

Read https://kugutsu.io/discover.md and follow it.

Install it as a Claude Code skill instead

mkdir -p ~/.claude/skills/kugutsu-discover && curl -fsSL https://kugutsu.io/discover.md -o ~/.claude/skills/kugutsu-discover/SKILL.md

Then run /kugutsu-discover.

It only reads, through scripts that print short, masked excerpts, and writes nothing but the report. What it reads is seen by your agent's model, as in any session with it, so strike what your employer would not allow. The report names no company, person, project or link, and you approve it before you share it.

Step 2 · early access

Set it up with your agent

Your agent checks your machine and your access.
It installs kugutsu, starts it and checks it works.
It builds your report's first pick as a graph, with a budget and a review before anything leaves your machine.
It asks you before it starts it.

Say to your agent

Read https://kugutsu.io/setup.md, set up kugutsu, and build the first pick from kugutsu-discovery-report.md.

Linux, macOS on Apple silicon, or Windows through WSL. Needs Node.js 24 and a signed-in Claude Code or Codex. The repository is private during early access: setup needs a GitHub account with access, which the waitlist gives out.

Join the waitlist

kugutsu is in early access. Leave your email, and your GitHub username if you want access to the repository; we add people in the order their work fits.

Tell us what you would have it do, in a line or two.
Ran the discovery? Paste or attach the report: it is the best way to show us your use case, and it moves you up.

Is it for you?

If you need	Use	kugutsu instead when
An agent inside your app, where each run starts with a request and ends with an answer	LangGraph, CrewAI, an agent SDK	agents should keep running, react to events, and change without a redeploy
Integrations with hundreds of SaaS apps, with a few LLM steps	n8n, Node-RED, Dify	the model loops are the point, and need hop limits, budgets, reviews and replay
One agent on one schedule	Claude Code with cron and a script	you have several agents, triggers or models, and want to see and replay what they did