oh-my-kimi (omk)

Local, proof-driven autonomous engineering runtime powered by Kimi. Beta MVP (0.4.x, pre-1.0).

Install · Quick Start · Docs

What is OMK?

OMK is a terminal-native autonomous agent that turns one high-level goal into planned, verified, and delivered repository changes. It runs entirely locally, requires no cloud control plane, and produces a proof.json artifact for every outcome.

• Chat-first surface — run omk to open a REPL. The classifier automatically escalates from quick answers through small edits to full goals. A visible engine pane shows what is happening under the hood.
• Headless when you need it — omk goal run "..." --until-ready drives planning, execution, verification, and review without blocking your terminal.
• Proof-backed delivery — every large goal decomposes into slices, passes a review wall, and lands as a PR with evidence. ready, not_ready, or blocked are the only terminal states.

Install

curl -fsSL https://raw.githubusercontent.com/ekhodzitsky/oh-my-kimi/master/install.sh | bash

Or build from source (Rust 1.78+):

git clone https://github.com/ekhodzitsky/oh-my-kimi.git && cd oh-my-kimi
cargo build --release

Also available via Homebrew and AUR.

macOS (arm64 / x86_64) and Linux x86_64. Windows is not supported yet.

Quick Start

omk setup                 # create config, state, and data directories
omk doctor                # verify Kimi CLI and environment

omk                       # open the chat REPL
# type a request and watch the engine pane (Tab to expand)

Run a goal headless:

omk goal run "Add OAuth and rate-limiting to the API" --until-ready
omk goal proof latest
omk goal replay latest

How it works

Single command surface. omk with no arguments opens a chat REPL. There is no separate "agent mode" to learn. The same session handles quick questions, file edits, and large goals.

Autonomous escalation. A classifier routes every request to the right backend: trivial answers, small edits, medium plans, or large goals. You see progress in the engine pane and can expand it with Tab. The default autonomous mode does not block on confirmation dialogs.

Proof-backed delivery. Large goals become durable tasks with planning artifacts, verification gates, and bounded agent waves. The controller runs a review wall (architect, code, test, security, performance, anti-slop), records an honest ready or not_ready verdict, and writes proof.json as the artifact. Slices run in isolated git worktrees with conflict detection.

What you can do

Goal execution flags

omk goal run accepts flags that control budgets, agents, and delivery:

Positioning

OMK is inspired by oh-my-claudecode and market-informed by the broader agentic coding landscape, including Devin, OpenHands, Claude Code, Aider, Dify, and Cody. It competes on durable goal state, explicit verification gates, and proof artifacts — not on feature breadth. See docs/COMPETITIVE_POSITIONING.md for the full market map.

Configuration

• Set KIMI_API_KEY in your environment.
• Config directory: ~/.config/omk/ (or $XDG_CONFIG_HOME/omk/).
• State directory: ~/.local/state/omk/ (or $XDG_STATE_HOME/omk/).
• Manage settings with omk config show, validate, and set.

Roadmap & docs

• ROADMAP.md — where we are headed
• SPEC.md — product spec and delivery contract
• AGENTS.md — multi-agent workflow and contributing rules
• docs/TUTORIAL.md — step-by-step first run
• docs/ARCHITECTURE.md — system design
• docs/API.md — machine-readable CLI outputs

License

MIT © oh-my-kimi contributors