Kimi writes.
OMK coordinates, verifies,
remembers, and guards.
Production-ready multi-agent orchestration harness for Kimi Code CLI (K2.6). Not a prompt pack. Not a model buffet. A Kimi-native control plane for shipping code with verification.
$npm install -g @oh-my-kimi/cli$omk init$omk doctor$omk chatWhy OMK?
OMK solves the six hardest problems in AI-assisted coding: premature completion, context corruption, memory loss, operator blindness, model chaos, and runtime drift.
Evidence Gates
Agents can't claim 'done' without proof. Require files, diffs, summaries, or passing commands before completion is accepted.
Worktree Isolation
Parallel workers run in isolated Git worktrees. Experiments stay reversible, and merge is always a deliberate step.
Local Graph Memory
Durable project memory as a graph: goals, decisions, risks, tasks, commands, files, evidence, and concepts.
HUD & Cockpit
Live operator visibility into run state, TODOs, ETA, usage, workers, and changed files. No more hidden agent state.
DAG Execution
Requests become task graphs with roles, dependencies, retries, fallback routing, and evidence requirements.
Decision Trace
Every policy decision is recorded in decisions.jsonl. Runs are inspectable and reproducible, never opaque.
How the Engine Works
OMK is built on 12 architectural pillars. Here are the four that define the runtime.
Kimi-native Control Plane
Built around Kimi Code. Project rules, agents, hooks, slash skills, and MCP configuration are shaped so Kimi remains the primary writer.
Evidence-gated Verification
Completion requires evidence: file exists, command passes, git diff is non-empty, or summary markers present. No narration-only completion.
Worktree-isolated Parallel Lanes
Parallel lanes run in isolated Git worktrees. Experiments are reversible, and review/merge is deliberate, not a side effect.
Persistent Graph Memory
Goals, decisions, risks, tasks, commands, files, evidence, and concepts stored as a graph. Smaller, safer context for Kimi.
Five Operating Rituals
OMK organizes daily work into five rituals. Each maps to a set of commands.
Ship
Implement with verification
omk chat, omk parallel, omk verify
Inspect
Run history & current state
omk runs, omk replay, omk inspect
Design
Visual/product direction
omk design, omk open-design
Remember
Durable project context
omk graph view, omk index
Guard
Safety & release confidence
omk doctor, npm run release:check
Quick Start
From zero to a verified run in six commands.
Installation
$npm install -g @oh-my-kimi/cli$omk --version$omk doctorRequirements
- Node.js 20+
- Git
- Kimi Code CLI installed and authenticated
- tmux for team/HUD workflows (Unix-like)
One Prompt to a Verified Run
$omk init$omk doctor$omk plan "Add a settings page with tests"$omk parallel "Implement the settings page"$omk verify --json$omk summary-showTip: OMK loads project rules, skills, hooks, MCP status, and Git state automatically. Evidence gates check required files, diffs, and summaries before marking completion.
CLI Command Map
Every command has an explicit maturity label. Stable commands are the normal operator path.
Start Here
Stable / Daily-use
Advanced
Alpha / Experimental
How OMK Differs
OMK is intentionally Kimi-first. Other models can advise, review, or QA, but the run remains bounded.
| Harness | Best When | Core Idea |
|---|---|---|
| OMC | Claude Code users | Team-first Claude Code orchestration |
| OMX | Codex CLI users | Codex workflow layer with reusable modes |
| OMO | Open multi-model routing | Open multi-model agent team with aggressive routing |
| OMK | Kimi Code with verification | Kimi-native DAG runtime with evidence gates and graph memory |
Ready to ship with verification?
Install OMK in 10 seconds and turn Kimi Code CLI into a bounded, verified coding team.
$npm install -g @oh-my-kimi/cli