OMK

Provider-neutral terminal coding agent with optional multi-agent orchestration, scoped tools, replayable sessions, and evidence-backed verification.

OMK supports interactive terminal use, non-interactive output, RPC integration, and an embeddable TypeScript SDK. It works with API-key and subscription providers without making one provider the control plane.

OMK targets state-of-the-art quality as a CLI coding-agent harness; it does not currently claim verified SOTA status. Comparative claims follow the same-model, reproducible protocol in Turn metrics and harness evaluation.

Install

Requires Node.js 22.19 or newer. Built-in local bash also requires sandbox-exec on macOS or bwrap plus unprivileged user namespaces on Linux. Enforcement is enabled by default and fails closed with sandbox.backend_missing; see Containerization.

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

Run without a global install:

npx --ignore-scripts open-multi-agent-kit

The published package name remains open-multi-agent-kit; the command is omk.

Quick Start

Start OMK in a project directory:

cd your-project
omk

Then:

1. Run /login to configure an OAuth subscription or API key.
2. Run /model to choose an available model.
3. Describe the change and the evidence required for completion.
4. Review tool calls, diffs, diagnostics, and test results before accepting the work.

OMK stores user configuration under ~/.omk/agent/ and project configuration under .omk/.

Optional document compiler

Install omk-book-to-skill to compile documents into reusable skills without adding Python extractors to OMK core:

omk install npm:[email protected]

It provides compile, update, and verification commands plus a local SHA-256 provenance manifest. See Book to Skill.

Providers & Models

OMK includes provider adapters for Anthropic, OpenAI Codex, Google, OpenCode, Kimi, Qwen, ZAI/GLM, native xAI/Grok, and other compatible services. Availability depends on the credentials and endpoints configured on the current machine.

• /login adds or selects credentials.
• /logout removes credentials for a selected provider.
• /model lists models available to the current credential set.
• Devin CLI subscription login exposes devin/swe-2 with medium, high, and max reasoning and a 1,000,000-token local context budget that selects the catalog's 1M-context lane. See Devin setup and verification limits and the Devin SWE-2 harness.
• NVIDIA NIM's z-ai/glm-5.2 entry transmits reasoning effort, including /think max.
• Native xai accepts subscription OAuth from /login or XAI_API_KEY for API billing. Only OAuth exposes weekly SuperGrok usage/reset. See Grok harness.
• The optional status rail shows independent provider quota windows when an official API or passive response signal is available, including Command Code 5-hour, weekly, and monthly credit meters.
• Missing quota values are shown as unavailable rather than estimated.

OAuth credentials are stored locally in ~/.omk/agent/auth.json. Multi-account providers keep accounts separate and send requests only through the explicitly selected account.

Quota and billing-cycle errors are classified as rate limits and can switch to the first authenticated resilience candidate before retry. See Provider resilience.

See Provider setup for provider-specific authentication, quota behavior, and environment variables.

Kimi For Coding

Configure Kimi through /login or the documented Kimi API-key environment variable. Requests use the fixed provider endpoint selected by the provider adapter.

Hugging Face

Set the documented Hugging Face token and select a compatible model. See the provider guide for current model and endpoint requirements.

Interactive Mode

Type / in the editor to open command completion.

| Command | Description | | --- | --- | | /login, /logout | Manage OAuth or API-key credentials | | /model | Select a model | | /think | Select a thinking level | | /settings | Change interactive settings | | /resume | Open a previous session | | /new | Start a new session | | /session | Show session path, messages, tokens, and cost | | /resource [probe\|policy] | Show resource pressure and effective concurrency for this run | | /debug [save] | Preview runtime diagnostics or explicitly save a metadata-only local report | | /goal [objective] | Show or set the durable goal; supports checkpoint <json>, pause, resume, evidence-gated complete, and clear | | /tree | Navigate the current session tree | | /fork | Create a new fork from a previous user message | | /clone | Duplicate the current session at the current position | | /compact [prompt] | Compact context, optionally with custom instructions | | /copy | Copy the last assistant message | | /name <name> | Set the session display name | | /scoped-models | Enable/disable models for Ctrl+P cycling | | /export [file] | Export the session | | /import <path.jsonl> | Import a JSONL session and continue it | | /share | Upload the session as a private GitHub gist | | /changelog | Show version history | | /reload | Reload project and user resources | | /hotkeys | Show keyboard shortcuts | | /star | Open the OMK GitHub repository | | /quit | Exit OMK |

Default Harness Safeguards

OMK warns from the third repeated identical tool call and blocks the sixth. It also repairs unmatched tool pairs before provider requests and adds guidance for supported Kimi, GLM, and Grok models. OMK_IDENTICAL_LOOP, OMK_TOOL_PAIR_REPAIR, and OMK_PROMPT_PRESET opt out; OMK_GOAL_CONTROLLER controls /goal and continuation.

See Usage and Keybindings for the complete interactive reference.

Repository Understanding

v0.97.0 introduced the OpenWiki workflow and policy, but its release artifact did not include the ignored corpus. The current interrupted corpus and partial checker remain worktree-only and blocked from trusted use pending exact source binding, output allowlisting, and pre-publication secret/private-path scans. .understand-anything/ remains optional local advisory data. See Runtime algorithms and direction.

Resource Governance

A built-in resource governor probes host capacity. Its default observe mode reports decisions without changing execution; opt-in adaptive and strict modes throttle tool and governed heavy-process concurrency. Inspect the current host with /resource or omk doctor resources [--json], and aggregate bounded local observations with omk doctor resources --report [--json]; see Settings.

Terminal Notifications

Interactive TTY sessions notify by sound after final prompt settlement by default. Successful prompts use a 5-second duration floor; failed and aborted/stopped outcomes notify immediately. Intermediate retries and continuations stay silent, and headless modes never play sound. Configure or disable this under notifications.completionSound or with OMK_COMPLETION_SOUND=0.

Sessions

Sessions are append-only JSONL transcripts stored under ~/.omk/agent/sessions/. OMK can resume, branch, export, and repair sessions without sending the transcript to a separate orchestration service.

Do not publish session files until you have reviewed them for source code, credentials, personal information, and private tool output.

Scripted Session Inspection

omk sdk session status --json
omk sdk session tail [id] --limit 20
omk sdk session inspect [id]
omk sdk session send <id> "message"

These commands inspect or append to stored transcripts. send does not wake or execute an agent.

Context Compaction

Automatic and manual compaction reduce older context while retaining recent conversation state. The current default compactor deterministically carries explicitly marked user rules outside LLM rewriting; custom hooks and unmarked prose remain unchanged. Cancellation restores queued user input without committing a partial summary.

See Compaction.

Configuration

Context Files

Run omk init --global --dry-run to preview portable user instructions, then omk init --global to create missing files without overwriting existing context or settings. See User context setup.

OMK discovers project instructions such as AGENTS.md from the current directory and parent directories. Use --no-context-files when a clean session must not load project context.

Prompt Caching

Provider prompt caching follows provider capabilities and configured retention. Cache metadata is local and bounded; provider-side retention remains subject to the provider's policy.

Environment Variables

Credential, network, model, and runtime variables are documented in Environment variables. Prefer the documented fixed provider endpoints for authenticated requests.

CLI Reference

omk [options] [prompt]
omk --help

Common modes:

omk                         # interactive TUI
omk -p "summarize changes"  # print mode
omk --mode json "task"      # structured event stream
omk --rpc                    # JSONL RPC mode
omk doctor resources [--json] # current pressure and admission report
omk doctor resources --report [--json] # bounded local observation aggregate
omk stats [--dir <path>] [--json] # aggregate turn metrics for a project

Skills Configuration

Skills load on demand from project and user directories. The global-only defaultActiveSkills setting keeps selected, user-scoped skills active in every prompt while their full instructions stay read-on-use. See Skills.

Other Options

Use omk --help for the authoritative option list for the installed version. Avoid copying options from an unreleased branch into automation.

RPC Mode

RPC mode reads JSON commands from standard input and emits JSON events to standard output. Use it when another process owns the UI or lifecycle.

See RPC.

SDK and Extensions

The TypeScript SDK exposes session creation, provider configuration, tools, event streams, durable-goal reducers, Goal/Core/Verified/Open/Next seam checkpoints, an explicit pass-gated advisory best-of-N judge, and policy helpers for compaction, retries, prompt budgets, cache transitions, and system-prompt assembly. Extensions can add commands, tools, UI components, themes, and resource loaders without patching OMK core.

omk-protocol provides the versioned TaskSpec -> ExecutionAttempt -> Observation -> EvaluationResult -> RuntimeDecision contracts and pure semantic reducers. See Run Protocol v1 for the implemented scope and migration boundaries.

• SDK
• Extensions
• Skills
• Themes

OMK + AdaptOrch

OMK is a local, MIT-licensed coding agent. AdaptOrch is an optional, separate evidence and risk layer for changes an AI wrote: it runs a change in an isolated copy of the project and reports what it touched, whether it ran, and what the tests said before and after. Its free Starter tier is self-hosted and runs on your own machine; a bring-your-own model key is required on every tier.

Review AdaptOrch plans → · claim boundary

For signup, plan comparison, or team/private-runner contact links without an API key or network request, run omk doctor adaptorch --links (add --json for structured output). See the onboarding guide. The command only prints links; it never creates an account, uploads code, or starts a run.

AdaptOrch is a separate proprietary product and is not bundled with OMK. Its published claim boundary states that it does not prove semantic correctness (correctness_claim=false); it reports and never rewrites, selects, or merges on your behalf. OMK does not require it.

Security and Privacy

• Tool access is explicit and can be scoped by path or runtime policy.
• Default OS isolation covers AgentSession's built-in local bash only; it does not isolate other tools, extensions, custom code, or the OMK process.
• Default tests scrub provider credentials; live provider tests require LIVE_E2E=1.
• Subscription quota requests use fixed HTTPS provider origins.
• Passive quota observers do not block model streams.
• Logs and issue reports must not contain API keys, OAuth tokens, cookies, or full private transcripts.

Report vulnerabilities privately through the repository's GitHub security contact rather than a public issue.

Development

git clone https://github.com/dmae97/omk.git
cd omk
npm ci --ignore-scripts
npm run build
npm run check
npm test

See Development and CONTRIBUTING.md.

Release notes: v1.2.0.

License

MIT. See LICENSE.