TokenRhythm's open-source agent harness and the product its later models come from. It is a microkernel agent with a CLI, a web UI and chat-channel front ends that all share one turn loop. It has persistent memory, a layered sandbox, built-in web search and on-device embeddings, and connects to 20+ model providers, including the TokenRhythm API, OpenRouter, Anthropic, OpenAI, DeepSeek and Qwen. The first release (v0.1.0rc1) came on May 12, 2026; v0.5.8 shipped on November 6. It is Apache 2.0 and has about 7,100 GitHub stars.

Its distinguishing piece is SquillaRouter, a local LightGBM and ONNX classifier. It scores each turn on length, language, code, keywords and embeddings, and sends the turn to the cheapest of four model tiers that can handle it. The July 2026 report "Agentic Routing: The Harness-Native Data Flywheel" makes the case that routing should be a harness-level decision. It conditions each choice on the full execution state rather than a single prompt, and chooses either one cost-effective model or a complementary ensemble. Every routing decision then becomes a labeled record whose label comes from the environment: query, harness state, model choice, trace, outcome and cost. That turns ordinary agent traffic into training data for better routers and models. This is the flywheel NeoHorse-1 was then trained on. The README says multi-model ensemble routing surpasses Claude Fable 5.

Self-reported comparisons, with costs from one shared price list:

  • PinchBench: forced single-model OpenSquilla on Opus 4.7 scores 93.85 against OpenClaw's 92.55, at 22% lower cost. A routed pool of DeepSeek-V4 Flash, GLM-5.1 and Opus 4.7 keeps 92.51 for $0.69 against $6.23.
  • ClawMark (100 business tasks): the harness adds 5.8 points to GLM-5.1, at higher cost.

The team also released claw-swe-bench, an adapter framework for running such harnesses on SWE-bench-style tasks.

agentsharnessroutingopen-source

Related