Open-source Rust infrastructure for AI agents. Our own agents run on it in production before anyone else uses it.
AI agents invent facts. They forget what they learned last session. They cannot tell a verified fact from a guess. One wrong answer is bad. Worse is that nobody can say where the answer came from.
We make infrastructure that keeps agents honest.
- eruka gives agents memory with confidence states. Each fact is confirmed, inferred, stale, or missing. The agent sees the gaps before it answers.
- ares runs the agents. It routes across model providers, tracks usage per tenant, and speaks the OpenAI API.
- thulp connects agents to tools and multi-step workflows. docs
- daedra gives agents web search with automatic fallback across 13 backends. docs
- deagle makes a codebase searchable for an agent. docs
- pawan is our coding agent. It uses all of the above. docs
- eruka-mcp: eruka in Claude, Cursor, and VS Code. crates.io · docs
- openeruka: a self-hosted memory server compatible with eruka. One binary.
- thulpoff: teaches small models from large ones with reusable skill files.
- dstack: persistent memory and quality gates for multi-repo agent work. crates.io · docs
- dwasm: builds Leptos WASM frontends. crates.io · docs
- dui: accessible Leptos components. crates.io
- lancor: a llama.cpp toolkit. docs
- aegis: typed manifests that generate configuration for the stack.
- nimakai: measures NVIDIA NIM model latency. Written in Nim.
Everything above is open source. Clone it and run it yourself.
We also run it as a managed service. The orchestration layer stays managed because running it safely is the product. The open parts are the same code we run.
- Each fix lands with a test that fails without it.
- Benchmarks live in the repos with the command and the date. Anyone can rerun them.
- Small crates that do one thing.
- Our agents do real work on this stack. A bug here breaks our own work first.
