kishu.
Kishu is the workflow layer for your AI coding agent. Route hard decisions to a thinking model, give the build to an efficient one, and get clear, checked work back.
One workflow. Your models. Less work to manage.
Set the boundaries.
Resolve the hard parts.
Take a focused brief.
Turn decisions into code.
One prompt connects task routing, model roles, and design guidance. Kishu works inside Codex, Cursor, and Claude Code.
Describe what you want to make or fix.
Choose the approach and the roles your task needs.
pstack playbooks.
Your models for each role.
Impeccable guides interface work, when needed.
What changed, what was checked, and what’s left.
A route, not a ritual. Small tasks stay small. Specialists and extra agents join when they help.
Explore AI coding workflows for a new feature, a tricky bug, and a refactor. See each model’s assignment and the handoff that moves the work forward.
$kishu:kishu-modeAdd team workspaces with roles and email invites.
Select any step to read at your own pace.
For a team invitation feature, the planner defines roles and expiry rules. The implementer receives a focused brief, relevant files, and acceptance checks. Design guidance shapes the invitation states; review checks the result.
You describe the outcome once. Kishu loads the relevant guidance.
Illustrative scenarios, not live agent runs. Model profiles are examples; your host’s available models and saved preferences determine the actual choices.
A high-cost model doesn’t need to write every line. Put deeper reasoning on the tricky decisions, and let an efficient coding model carry the plan forward.
The mix can reduce cost and waiting. Actual savings depend on the task, model prices, context, and coordination overhead.
Try a different mix. The walkthrough updates with you.
Deeper thinking at the decision points. Efficient coding for the implementation.
A local preview of role preferences. Copy the request into your agent to choose supported models and reasoning levels with Kishu.
ADHD-aware communication puts the outcome first. Find the change, the proof, and the next decision without digging through a wall of process.
Bring this to your next buildKishuExample handbackThis site was built with Kishu mode: a routed brief, considered design, interactive examples, and checks against the actual result.
The proof is in the project.
Explore the source, the demo’s behavior tests, and the record of browser checks behind this page.
The site is real. The scenarios above are demonstrations of how the workflow can fit different tasks.
Three working demos. Three very different briefs.
Open one and try it for yourself.
Explore Veldrift
An immersive MMO concept
A world worth wandering through. Original generated artwork, cinematic type, and three landscapes you can explore at your own pace.
Try the dashboard
A simulated trading workspace
A dark, focused dashboard with moving prices, explorable charts, and a working demo order flow. Every price and trade is fictional.
Visit the studio
A creative workshop booking site
Warm photography, a little breathing room, and a clear path from curiosity to a class. Choose a workshop and try the local booking flow.
Real Kishu builds. Fictional products.
No live markets, payments, or reservations.
Install Kishu in your agent, choose the model roles that suit you, and bring the thing you’ve been meaning to build.
Explore Kishu on GitHub1. Clone the repo and install the plugin.
git clone --branch v0.1.5 --single-branch https://github.com/SPDUK/kishu-stack.git
cd kishu-stack
python scripts/kishu.py install --harness codex2. Start a new chat. Give Kishu a challenge.
$kishu:kishu-mode
Select Kishu through /skills, or use the command above. Python 3.11+ and Git are needed for installation.
The practical details: where it works, how you choose your models, and what the workflow adds.
Kishu is an open-source workflow plugin for AI coding agents. It brings task routing, customizable model roles, design guidance, and ADHD-aware communication together in one entrypoint: Kishu mode.
Kishu provides native plugin entrypoints for Codex, Cursor, and Claude Code, plus a standalone skill export. It works inside your existing agent with the code and tools available in that environment.
Yes. Kishu setup helps you choose supported models and reasoning effort for planning, implementation, review, and other roles. Saved preferences guide delegation when it is useful. Available models depend on your host; role preferences do not change the model running your current chat.
It can concentrate expensive reasoning on difficult decisions, give efficient coding models a focused brief, and let independent assignments run in parallel. Savings depend on the task, model prices, context, and coordination overhead. The examples on this site are illustrative, not benchmarks.
pstack provides the engineering playbooks behind Kishu’s routing and handoffs. Impeccable provides the foundation for its design guidance. Kishu adapts these foundations with its own preferences for scope, model choices, useful checks, and clear communication; i-have-adhd informs its communication style.
Kishu brings three open-source foundations together with its own working preferences: care where it counts, less process to manage, and your choices at the center.
Licenses & full attributionOriginal licenses and notices stay with the bundle. With thanks to the authors; attribution does not imply endorsement.