Gemma-first, OpenClaw-powered, community-informed.
If you have heard of OpenClaw or Hermes and want to understand where Gemmaclaw sits, this page explains the relationship. Short version: Gemmaclaw is a focused distribution built on top of OpenClaw, shaped by experience in both the OpenClaw and Hermes communities, and aimed squarely at users who want a Gemma-powered local agent that works on real hardware with minimal friction.
These are distinct efforts with different scopes and audiences. None is a replacement for the others.
OpenClaw is a general-purpose agent framework with multi-provider support, multi-channel delivery, a plugin architecture, and a broad runtime. It is the upstream project that Gemmaclaw is built on. Gemmaclaw pulls in OpenClaw updates directly and inherits its core runtime improvements.
Gemmaclaw is built on OpenClaw and targets the same npm ecosystem.
Hermes is another serious project in the local-agent ecosystem. It focuses on practical experiments with local tool-using models and has produced useful community knowledge about what local models can reliably do. Gemmaclaw learns from participating in the broader local-agent community that includes Hermes.
Gemmaclaw does not share code with Hermes, but the broader local-agent conversation that includes Hermes has informed Gemmaclaw's thinking about what matters for real users on real hardware.
Gemmaclaw is a Gemma-first distribution layered on OpenClaw. It adds hardware-aware setup, model and runtime recommendations, local-first self-hosting guides, a public benchmark suite, field notes from real usage, and runtime behavior enhancements for local agents.
The relationship is straightforward: OpenClaw provides the runtime foundation, and Gemmaclaw builds a Gemma-specific distribution on top of it. Community participation, including in the OpenClaw and Hermes communities, feeds back into Gemmaclaw's priorities and field notes.
If you want to run a Gemma-powered agent on your own hardware and need clear guidance on which model, which backend, and what to expect, Gemmaclaw is designed for that.
One command gets a working Gemma assistant. The setup wizard auto-detects your hardware and picks the right model size and quantization.
GPU, CPU-only, and cloud paths are all first-class. Benchmark results and field notes tell you what actually works at each hardware tier.
All models tested on the same 51-task agentic suite. Results include full transcripts, per-task scoring, and LLM judge evaluations. No hidden runs.
Practical setup guides for real users. llama.cpp first, Docker sandbox, CPU-only fallbacks, and reproducible configurations.
Code-owned instructions that help local agents verify external delivery receipts and follow through on commitments before claiming success.
Gemmaclaw pulls runtime improvements from OpenClaw continuously. You get the OpenClaw runtime foundation plus Gemma-specific additions.
Gemmaclaw is a volunteer-driven project and is not an official Google product. It uses Gemma models and follows Google's open-model terms, but is built and maintained independently. The Gemmaclaw community works alongside and learns from both the OpenClaw community and the broader local-agent ecosystem, including projects like Hermes.
If you are building on Gemmaclaw or want to contribute benchmarks, hardware reports, or enhancements, the Community page has field notes and community hardware reports, and the GitHub repository is the right place for contributions.
Get a Gemma agent running in minutes. The wizard auto-detects your hardware and picks the best model.
See how Gemma models perform on the agentic task suite. Full transcripts and per-task scoring included.
Learn about the runtime enhancements Gemmaclaw adds beyond the OpenClaw baseline.