Skip to content

Agentic delivery that stops at your gates.

Agents do the work. You hold the gates.

Bankai is an open-source, AI-driven development lifecycle for Claude Code, Codex, Cursor and Antigravity. Deterministic CLI ops, governed agents, per-stack handbooks.

Coming later: Bankai-Library and hosted, autonomous runs. The path keeps going.

claude code · a request to Ready
/kurapika Add a "Clear all" action to the notifications list
ren · breath → rasengan → kokusen → amaterasu → spiritual-message → jutaisho
committed locally · nothing pushed
/hatsu:aka
gyo → squash unpushed → ao → gyo → push
/hatsu:mukai
murasaki → hanten → kokusen → kotoamatsukami → byakugan → shibari → landing report
one PR opened · offering en
/hatsu:en on HA#12
sharingan → observe CI and the current-head round → Ready
G2 · the merge is yours

A real flow: work a request, push, open the PR, drive it to Ready. The merge stays yours.

Agents are fast; lifecycles are messy.

PRs, releases and backlogs drift when every session improvises.

How it works

Three live parts, one loop. Library comes later.

  • Nen

    deterministic ops

    v0.29.0

    A CLI for the mechanical work agents should not improvise.

    github.com/zheref/nen
  • Hatsu

    governance harness

    v0.159.0

    A skill and agent plugin with 60+ skills, 17 agents and human gates G1-G5.

    github.com/zheref/hatsu
  • Bankai-Handbooks

    the canon

    v0.13

    Per-stack practices for SwiftUI+TCA, Jetpack Compose and React+Redux.

    github.com/zheref/bankai-handbooks
  • Bankai-Library

    coming later

    Coming later

    Reusable components, next on the path. Coming later.

  • Cloud tier

    coming later

    Coming later

    Hosted, autonomous runs. Coming later.

Every stage of the lifecycle, with a skill behind it

Lead with the plain-English verb, then the code name. Each card names the skills that run the stage and whether it is available today.

  1. 01Available

    Intake and spec

    A raw idea interviewed into a decision-complete brief, challenged, then filed as one issue.

  2. 02Available

    Build

    One request per turn in an isolated worktree: warm up, author, verify, commit locally. Never pushes.

  3. 03Available

    Review and test

    Adversarial review before the PR, the impacted suites, coverage against the ladder.

  4. 04Available

    PR to Ready

    Push when you decide, one PR with a composed body, then the readiness watch to your merge prompt.

  5. 05Available

    Backlog operations

    A severity band, a label, a hand-picked list or the whole surface-matched backlog, delivered to gates.

  6. 06Available

    Release

    The tag cut with its own proposal PR, a non-production send, and publication only on your typed go.

  7. 07Available

    Adopt and maintain

    Make a repository a consumer, keep its canon mirror current, raise runners, install a CI tier, update Hatsu.

  8. 08Available

    Validate and clean up

    Post-merge UI validation on your call, the leftovers inventoried and removed only as you picked.

The layers, and the five human gates

Nen at the base, Hatsu on top, the Handbooks as canon. The Cloud tier is the future autonomous plane; the Library is roadmap. Select a layer or a gate.

Human gates

Agents propose; you decide. Hatsu drives work to a gate and stops there.

Hatsu

The governance harness: a plugin of skills and agents for Claude Code, Codex, Cursor and Antigravity. It drives work from a request to a human gate; every deterministic step is a Nen verb.

Quick start

One checkout serves every surface. Pick yours: the install command, the session command, a first request, and the Nen bootstrap at the version Hatsu pins.

Install
git clone https://github.com/zheref/hatsu.git ~/Code/Agents/hatsu     # or wherever you keep it
~/Code/Agents/hatsu/scripts/hatsu_surface_link.sh --surface claude-code --root ~/Code/Agents/hatsu

Open a session

/kurapika <request>

A first request

the next message, or /hatsu:ren <request>
A session, end to end
/kurapika Add a "Clear all" action to the notifications list, with a confirmation sheet
Full guide: docs/using/claude-code.md

Nen bootstrap v0.28.0

Hatsu's warm-up installs this pin for you, checksum-verified. The command below is the same bootstrap by hand, at the ref Hatsu's contract pins.

hatsu nen/contract.json · dependency.minimum 0.28 · pinned_ref v0.28.0 · zheref/nen

nen bootstrap · v0.28.0
curl -fsSL https://raw.githubusercontent.com/zheref/nen/v0.28.0/bootstrap/nen.sh -o nen-bootstrap.sh
bash nen-bootstrap.sh --ref v0.28.0

Open source, complete on its own

MIT, fully public and free, built in the open by zheref. Everything below the line runs on your own credentials today.

Open source versus the Cloud tier
Open source · MITCloud tier, coming later Coming later
Nen, the deterministic CLI● Included● Included
Hatsu, the governance harness● Included● Included
Bankai-Handbooks, the canon● Included● Included
Bankai-Library (coming later)◌ Coming later◌ Coming later
Runs on your own credentials● Included● Included
Autonomous runs in your org— Not included◌ Coming later
GitHub App agents— Not included◌ Coming later
Managed or self-hosted runners— Not included◌ Coming later
Always bring your own model API key● Included● Included

The open-source system is complete on its own: nothing here is crippled to sell a hosted tier.

The maintainer's own Claude subscriptions are used only for building Bankai itself, never for customers' work. Customers always bring their own model API key.

Hosted, autonomous runs are coming later. Get notified.

A waitlist opens later.

Support the project

Bankai is built in the open by one maintainer. Sponsorship keeps it that way.

What sponsorship funds

  • Maintaining Nen, Hatsu and the Handbooks across four agent surfaces
  • Model usage for building and testing Bankai itself
  • Documentation, examples and the public site
Sponsor on GitHub

Sponsors

No sponsors yet. The first names appear here.

Governed AI delivery setup for dev teams

Delivered by ValleIA, a product that leverages Bankai. Bankai itself stays open source and free; this is the hands-on setup for a team that wants it installed and tuned.

Fixed-fee setup

Install the CLI and the harness in your repositories, tune the handbooks for your stack, and train the team on the gates.

From$2,500– $7,500TBD

Care plan

Optional monthly care: updates, repins and a standing line to the people who set it up.

From$500– $1,500per monthTBD

Contact usValleIA · TBD

FAQ

Is it free?

Yes. Nen, Hatsu and the Handbooks are MIT licensed, fully open source, public and free. A hosted tier is coming later and is not for sale yet.

Which agents and models?

Hatsu runs inside Claude Code, Codex, Cursor and Antigravity. A tier is named, never a model version: frontier for your own session, deep for reviewers, fast for workers, economy for watchers. The live model id is read from your surface when a skill runs.

Does it merge or publish on its own?

No. Every PR ends at your merge prompt, and publication runs only on your typed go. Agents propose; you decide at the gates.

Where does my code run?

On your machine, in the agent you already use, on your own credentials. There is no hosted Bankai today.

Who pays for model usage?

You do, with your own key or subscription on the surface you run. The future Cloud tier will also require your own model API key.

Which stacks?

The Handbooks cover SwiftUI+TCA, Jetpack Compose and React+Redux Toolkit, plus baselines for security, UX, release and quality that apply everywhere.

How do I update?

Run the update skill on your surface: it fast-forwards the checkout you keep and ends with the one thing left to type.

Is there a hosted version?

Not yet. Hosted, autonomous runs are coming later, unnamed and unpriced until then.

Built in the open by zheref. MIT license.

Versions and commands on this site are read from the repositories at build time. hatsu 0.159.0 · nen 0.29.0 · handbooks v0.13 · 2026-11-04

Independent open-source project. Not affiliated with or endorsed by any anime or manga rights holder.