What we build
One foundation, and the things that run on it: a library, a harness, a research fork, and the models we publish from it.
Agention
A TypeScript-native library for building AI agents, built on the idea that most agent frameworks trade away control for convenience. Agention gives you enough structure to be productive and enough transparency to stay in control.
- 01Multi-provider: Claude, OpenAI, Gemini, Mistral, and local models via Ollama and llama.cpp behind one interface.
- 02Composable: agents are objects, pipelines are arrays. Sequential, parallel, or voting workflows, no hidden abstractions.
- 03Observable: per-call token counts, timing, and pipeline visualization built in.
- 04Multimodal: a unified message-content interface for images across every provider.
- 05Context management: provider-agnostic persistent history, with plugins for compression and result masking.
- 06Delegation: wrap a sub-agent as a tool for hierarchical workflows.
Marshall
A CLI harness, built on Agention, for orchestrating multiple models on a single coding task. Name an agent, pin a model, give it a job: a planner, a coder, a reviewer, each running whichever model actually suits the role, local or hosted.
- 01Multi-agent: assemble teams of specialized agents instead of relying on one model for everything.
- 02Local-first: llama.cpp and Ollama are first-class, no API key or cloud dependency required.
- 03Three control levels: approve every action, run hands-off, or let the model judge which tool calls need sign-off.
- 04Split across providers: mix local inference with OpenRouter, Claude, OpenAI, Gemini, and Mistral in the same run.
- 05Steerable: interrupt and redirect mid-session, with project memory carried in
AGENTS.md.
Bricks
A lab experiment: a browser game built end to end by Marshall, running Qwen3.8-27B. A brick island floats in the clouds under zombie siege: score, health, ammo, a pointer-lock aim mode, orbit camera, and a day/night cycle, all shipped by the harness from a brief.
Play Bricks ↗Quantized Models
Requants of Qwen3.8-Flash-Next, published on HuggingFace: one for anyone running mainline llama.cpp, one built specifically for AMD Strix Halo's unified memory using our ROCmFPx format.
See all builds →