OMK — Freedomd agent control plane

Local-first, multi-provider, evidence-based orchestration for coding agents.

OMK turns a coding goal into a scoped DAG run, routes it across compatible runtimes, limits authority per lane, records local evidence, and keeps the loop moving even when a provider is unavailable, retained, export-restricted, or policy-blocked.

What changed in v0.80.2

OMK now has a Freedomd control-plane direction:

• Provider sovereignty scoring — route by health, authority, capability, retention, jurisdiction, cutoff risk, locality, and portability.
• Retention gate — classify task data and decide allow, redact, downgrade, or block before a prompt leaves the local control plane.
• Cutoff-resilient routing — provider incidents and policy blocks degrade to local authority, read-only review, artifact replay, or a clear diagnostic instead of silently stalling.
• Local-first evidence envelopes — persist provider-independent evidence under .omk/runs/<run-id>/freedomd/ with hashes for replay and audit.
• Default agent freedom — high-autonomy agent behavior is default; strict guardrails are opt-in with OMK_STRICT_GUARDRAIL=1.
• Explicit Freedomd mode — sovereignty routing is opt-in with OMK_FREEDOMD_MODE=balanced or OMK_FREEDOMD_MODE=strict.

30-second demo

npx -y -p open-multi-agent-kit omk do "review this repo for release risk" --dry-run --json

A dry run produces bounded planning artifacts before changing files:

goal
  review this repo for release risk

flow
  goal -> input envelope -> DAG -> scoped lanes -> evidence -> verify

artifacts
  .omk/runs/<run-id>/input-envelope.json
  .omk/runs/<run-id>/dag.json
  .omk/runs/<run-id>/dag-compile-report.json

Executing a run adds loop state, route decisions, evidence observations, and replayable artifacts under .omk/runs/<run-id>/.

Install

npm install -g open-multi-agent-kit
omk init
omk doctor
omk chat

Local checkout verification:

npm ci
npm run build
node dist/cli.js do "explain this repo" --dry-run --json
npm run verify:no-kimi

Configure Freedomd

Balanced mode prefers provider diversity and records sovereignty diagnostics:

OMK_FREEDOMD_MODE=balanced omk chat

Strict mode prefers local/self-hosted paths and blocks retained/export-restricted providers more aggressively:

OMK_FREEDOMD_MODE=strict OMK_MAX_RETENTION_DAYS=0 omk chat

Provider incident overrides are local and auditable:

{
  "schemaVersion": "omk.provider-incidents.v1",
  "incidents": [
    {
      "providerId": "anthropic",
      "kind": "export-control",
      "severity": "block",
      "reason": "operator policy override",
      "updatedAt": "2026-06-16T00:00:00.000Z"
    }
  ]
}

Save that as .omk/provider-incidents.json, or set OMK_PROVIDER_INCIDENTS to the same JSON. Optional public feeds can be supplied with OMK_FREEDOMD_FEED_URLS; feed failures never block routing and local overrides win.

Use OMK if you

• use Codex, OpenCode, CommandCode, Kimi, DeepSeek, GLM, Qwen, OpenRouter, MiMo, or local models in the same repo;
• want provider availability, retention, jurisdiction, and exportability treated as runtime risks;
• need read/write/shell/merge authority separated by task;
• want agents to produce replayable evidence before saying “done”;
• need project-scoped MCP, skills, hooks, and graph memory instead of ambient global tool access;
• need degraded local/review/replay modes when frontier providers are blocked.

What OMK controls

Goal -> DAG -> lane grants -> runtime routing -> evidence envelope -> verify -> replay

• Routing: choose compatible provider/runtime lanes with capability and sovereignty scores.
• Authority: keep read, write, patch, shell, MCP, and merge power explicit.
• Evidence: require command output, diff, artifact, metric, or review proof.
• Replay: save run artifacts under .omk/runs/<run-id>/ for audit.
• Fallback: degrade to local/review/replay modes when policy or provider health blocks execution.
• Scope: keep MCP servers, skills, hooks, and memory bounded per lane.

How OMK differs

| Surface | Common failure | OMK adds | | --- | --- | --- | | Single-agent coding CLIs | Strong execution but weak audit | DAG, lane grants, local evidence, replay | | Cloud frontier providers | Policy/retention/export cutoff risk | Freedomd sovereignty scoring and degraded fallback | | MCP servers | Ambient tool access | Per-lane MCP/tool/shell authority | | CI | Post-hoc verification | Agent-time evidence gates before completion | | Local LLMs | Lower capability but high sovereignty | Local-first fallback and review/replay modes |

Release and safety claims

Current source version: [email protected].

• Runtime contract family: v1.2.
• Release channel: pre-1.0.
• OS-level sandboxing is not claimed; safety relies on authority gates, approval/sandbox policy, scoped capabilities, evidence, and local artifacts.
• npm latest claims are valid only after tagged CI and registry verification both pass.
• README hero generation used the local Codex OAuth gpt-image-2 workflow; the official omk image OpenAI Images API path still requires an OpenAI Platform project API key supplied only at runtime.

See versioning, claims, provider maturity, and SECURITY.md.

Examples

• Codex MCP evidence run: project-scoped MCP setup plus evidence-gated DAG dry run.
• Provider fallback: --provider auto routing with fallback planning.
• Proof index: source-controlled proof bundles checked by npm run proof:check.

Community

• Ask setup and provider questions in Discussions.
• File reproducible bugs in Issues.
• Star the repo to follow the 0.80.x pre-1.0 Freedomd release line.

License

MIT