Secure runtime for autonomous agents

The model proposes.
The runtime disposes.

Gateway • Policy guardrails • OpenShell sandbox • Durable skills (Hermes)

Zero hard dependencies • Pure stdlib core • MIT licensed

GATEWAY

Single control point

Routes inference and adjudicates every proposed action before it reaches the sandbox. Explainable decisions on every step.

POLICY + SANDBOX

What actually runs

JSON policies (shell/fs/net) that are easy to review. Sandbox executes only allowed commands with timeout + full I/O capture.

HERMES + SKILLS

Gets better with use

Captures proven goal→plan runs as durable, replayable skills. Next matching intent replays instantly.

Get started

Run the reference runtime locally

git clone https://github.com/cobusgreyling/nemoclaw-runtime.git
cd nemoclaw-runtime
./nemoclaw.sh
nemoclaw onboard
nemoclaw run "check system disk health"
nemoclaw run "clean up everything on the host"   # watch blocks
nemohermes run "audit inventory"                 # captures skill
nemoclaw serve                                   # → http://127.0.0.1:8848

See docs/quickstart.md, BLOG.md, and the architecture docs.