OMK
GitHub Install

open-multi-agent-kit · v0.95.1

Software engineering
with receipts.

OMK is a provider-neutral coding agent and multi-agent control plane. Turn a goal into bounded DAG lanes, block “done” without fresh evidence, and keep every run replayable — for Codex, Claude Code, OpenCode, and local models.

Get started npm package

Node ≥ 22.19 MIT npm install -g open-multi-agent-kit

omk · CONTROL
evidence gate · PASS
goal

Ship harness-graph health gate with CI and scorecard.

L1 · graphdone
build-harness-graph.mjs

1421 nodes · 2065 edges

L2 · analyzedone
graph_analyze.py

bipartite SPOF · communities

L3 · gateverify
health_gate.py --json

waiting fresh evidence

L4 · docsrun
SCORECARD.md

96/100 · write receipt

terminal · lane L3
$ python3 .omk/harness-graph/health_gate.py --json
{ "status": "PASS", "dead_skills": 0, "dead_hooks": 0
$ npm run check
✓ release surface OK · shrinkwrap OK · tsgo clean

Routes across models without rewriting the control plane.

  • Codex
  • Claude Code
  • OpenCode
  • Kimi · GLM · local

Product

Built like a coding product,
not a chat wrapper.

The same shape as a modern agent IDE home: clear job-to-be-done, a live workspace, and short paths from install to first verified run.

Parallel lanes

Own paths. No stomping.

Goals become a DAG. Each lane gets resource claims and write ownership so agents stop overwriting each other mid-flight.

OMK parallel lanes demo

Evidence gate

Green means measured.

Acceptance predicates block completion until commands leave fresh receipts. Chat claims stay labeled unverified.

Provider-neutral

Swap models, keep control.

Codex today, Claude tomorrow, local for air-gap. The evidence model does not change with the vendor.

Session recovery

Crash mid-run? Replay.

Durable session state and repair keep long orchestration jobs recoverable instead of restarting from zero.

OMK session doctor demo

How it works

Scope. Verify. Replay.

  1. 01

    State the goal

    OMK decomposes work into a bounded DAG with owned paths and stop predicates.

  2. 02

    Run parallel lanes

    Skills, hooks, and MCP tools route per lane. One writer per mutable surface.

  3. 03

    Gate on evidence

    Completion waits on fresh command output, tests, and health checks — not vibes.

  4. 04

    Keep the receipt

    Replayable artifacts stay available for review, recovery, and release notes.

OMK//CONTROL

A control plane for multi-agent work.

Interactive coding agent CLI underneath. Orchestration, skill routing, MCP health, and harness-graph inventory on top — one operator surface.

  • Skill hubs (including marketing/SEO via omk-marketing)
  • Harness graph: agents × skills × hooks × MCP
  • Fail-closed health gate and CI workflow
  • Context budget and session doctor
OMK TUI control plane screenshot

Install

One global binary.

Same path as a Codex-style product page: install, boot, first goal. 

npm install -g open-multi-agent-kit@0.95.1 --ignore-scripts
omk --version
omk

Docs & FAQ

Straight answers.

Is OMK a coding agent or an orchestrator?

Both. open-multi-agent-kit is an interactive coding-agent CLI. OMK//CONTROL adds multi-agent DAG lanes, skill/MCP routing, and evidence gates.

Which models are supported?

Provider-neutral via omk-ai: Codex, Claude, OpenCode, Kimi, GLM/ZAI, local providers, and more. Swap models without changing the control model.

How is completion verified?

Acceptance predicates plus fresh command evidence. Unverified runs stay labeled UNVERIFIED. Green chat is not a release signal.

Where do I read more?

GitHub · npm · docs index · v0.95.1 notes