OmniForge

OmniForge is a local-first, open-source AI automation app:

you describe it, the LLM builds it, you run it.

No YAML. No marketplace. No config sprawl.

Built for users first

OmniForge is for people who want automation outcomes, not framework plumbing.

• Describe what you want in plain language.
• Let OmniForge generate and run the workflow.
• Keep control of your data and provider keys locally.

What ships

• End-user CLI app (omniforge) for onboarding, creating automations, and running agents
• Local-first runtime with encrypted secrets and portable state in ~/.omniforge/
• Provider support through providers/catalog.json
• Reusable skills in skills/
• Open-source monorepo for transparent, community-driven development

Quick start (recommended)

Install the CLI globally:

npm install -g omniforge

Run onboarding:

omniforge onboard

Update configuration later (provider/model + web search):

omniforge config

Create your first automation:

omniforge create "Monitor my Gmail for investor replies and draft follow-ups in my tone"

Resume an existing agent:

omniforge agents

Run from source (development)

git clone https://github.com/siddakBath/omniforge.git
cd omniforge
npm install
npm run build

First-time onboarding:

npm run dev:cli -- onboard

Create and manage agents:

npm run dev:cli -- create "Your automation description"
npm run dev:cli -- agents

Optional: configure scheduler startup from source:

npm run dev:cli -- scheduler-service install

Architecture overview

See docs/architecture.md.

At runtime, each turn follows:

user message
     ↓
LLM call (system prompt + recent history + tools)
     ↓
if tool_use → execute tool → append tool_result → loop
if text     → stream to user → await next user message

Agents are checkpointed after complete turns and can be resumed.

Local storage model

OmniForge writes all state to ~/.omniforge/:

• config.json — generator defaults and provider keys
• config.json — generator defaults, provider keys, and web search settings
• params.json — required skill params (secret values encrypted at rest)
• skills/<skill-id>/SKILL.md — reusable generated skill playbooks
• agents/<agent-id>/agent.json — persisted agent state (messages, checkpoints, status)
• agents/<agent-id>/system-prompt.md — agent system instruction
• agents/<agent-id>/data/ — agent working directory for file tools

Optional per-agent daily schedule is stored in the schedule field inside agent.json.

Runtime context is bounded automatically to the most recent messages (default 60), while always injecting system-prompt.md at the top for every model call. Configure with OMNIFORGE_MAX_CONTEXT_MESSAGES.

Scheduling is timezone-aware (IANA timezone) and runs once per day at configured HH:mm local time. The scheduler dynamically reloads schedule changes and runs catch-up jobs missed during downtime.

Contributing

See CONTRIBUTING.md for setup, standards, and pull request requirements.

Open-source project standards

• License: LICENSE
• Security policy: SECURITY.md
• Code of conduct: CODE_OF_CONDUCT.md

Notes

• Provider catalog includes: Anthropic, OpenAI, Gemini, Ollama (local), DeepSeek, xAI, Groq, Mistral, OpenRouter.
• Anthropic and Gemini use native SDK paths; compatible providers use OpenAI-style API routing.
• Skills are markdown playbooks with YAML frontmatter metadata.
• Tools are runtime capabilities and are injected separately from skills.