🧭

Thymos

A Rust runtime that separates cognition from execution so LLM agents can't touch the world without a typed, auditable ledger entry.

The pipeline

cognition → Intent → compiler → Proposal → policy → ledger → world

Every step is typed. Every commit is signed. The cognition gateway never mutates state — it only proposes.

Drop-in providers

anthropic — Claude w/ native tool use + prompt caching
openai — OpenAI or any compatible endpoint
huggingface — HF Router, serverless OSS models
lmstudio — local LM Studio on :1234
local — Ollama, vLLM, llama.cpp
mock — deterministic test fixture

Quickstart (HF Router)

# .env
HF_TOKEN=hf_...
HF_MODEL=Qwen/Qwen2.5-Coder-32B-Instruct

cargo run -p thymos-server
thymos run "explain the ledger module" --provider huggingface