OneFactory is a local-substrate AI agent system: you give it a goal, and it builds, verifies, and ships — autonomously. No cloud. No external APIs. Air-gap capable.
The problem: today's AI agents need constant babysitting — prompts, retries, glue code. You ask for a result, you get a draft that needs three more rounds of your time.
The solution: a factory, not a chatbot. One API call spawns a task agent. It reasons about the plan, executes against your local models, then verifies its own work through a triple loop — write, verify, fix until clean. Nothing ships unverified. Failures feed back into the loop with concrete feedback until the output passes or the loop budget is spent — loudly, not silently.
Built for: indie hackers and teams running agents on their own hardware; anyone who can't — or won't — send data to the cloud.
Key features: spawn-agent API — one call, one verified result; build-repo — describe software in plain English, get a codebase; agent registry — every spawned agent tracked; MCP tool layer — filesystem, shell, GitHub, fetch, memory; safety layer on every run.
The AI inside: Llama 3.1 70B (scales to 405B) as the reasoning core, Llama-Guard safety, Model Context Protocol for tool use, recursive agent spawning. Demo video shows the same real factory code running against a smaller local model.
Impact: $0 cloud cost, zero data exfiltration, agents that finish jobs instead of producing drafts. Open source.
Built With
- fastapi
- llama-3.1
- llama-guard
- local-llms
- mcp
- ollama
- python
