HermesClaw is a community stack that runs the Hermes Agent (Nous Research) inside NVIDIA's OpenShell sandbox. OpenShell enforces limits on network egress, filesystem writes, and system calls at the kernel level, so the agent keeps its full tool set while the operating system, not the prompt, decides what it may touch. OpenShell intercepts every call to inference.local inside the sandbox and routes it to the configured backend.
The project can also run without OpenShell as a plain Docker container (multi-arch, linux/amd64 and linux/arm64) with a llama-server on the host, which is the path for macOS and machines without an NVIDIA GPU. The full sandbox mode requires Linux, an NVIDIA GPU, and an OpenShell install. Hermes's system prompt is around 11k tokens, so a 32k context window is needed for local models.
Enforcement in sandbox mode is layered: network egress through an OPA policy and HTTP CONNECT proxy to approved hosts only, filesystem access via Landlock limited to ~/.hermes/, /sandbox/, and /tmp/, seccomp BPF blocking ptrace, mount, kexecload, perfeventopen, and processvm_*, and a privacy router that strips credentials from the agent and injects backend credentials outside the process.
Features
- Policy presets: strict (inference only), gateway (adds Telegram and Discord), and permissive (adds web search and GitHub skills), switchable without a restart
- Hermes capabilities retained: persistent MEMORY.md and USER.md, self-improving skills (DSPy + GEPA), 40+ built-in tools, cron tasks, multi-agent delegation, MCP servers, and IDE integration over ACP (VS Code, JetBrains, Zed)
- Messaging: Telegram and Discord under the gateway policy; Signal, Slack, WhatsApp, and email under permissive; voice notes are transcribed before reaching the model
- Inference backends: llama.cpp, NVIDIA NIM, OpenAI, Anthropic, Ollama, and vLLM
- Skills library: installable skills for a weekly arXiv digest, code review, database anomaly detection, market alerts, a Slack support bot, and Home Assistant control
- Use-case guides: seven end-to-end setups (researcher, developer, home automation, data analyst, small business, privacy-regulated, trader) each with automated tests
- CLI: hermesclaw onboard, start, stop, status, connect, logs, policy-set, doctor, and chat
- Personalisation: a persona.yaml loaded into every session and ~/.hermes/SOUL.md injected into the system prompt
- Comparison with NemoClaw: a maintained feature table and test results contrast HermesClaw with NVIDIA's OpenClaw-based reference stack
