Skip to content
UsecaseBoard
← All use cases
githubsecurity

ClawShell isolates OpenClaw and Hermes API keys behind a privileged proxy

At a glance

Runta built ClawShell so OpenClaw/Hermes never hold real LLM API keys—virtual keys map to real ones in a privileged process with DLP scanning and email sender filters. This workflow is listed under security, credentials, sandboxing, openclaw, hermes, dlp.

Agents / tools
ClawShell, OpenClaw, Hermes Agent

Building a tool for these workflows? See weekly rates.

What they did

After prompt-injection experiments exfiltrated session tokens from a standard OpenClaw setup within minutes, they open-sourced ClawShell: a Rust sidecar that sits between OpenClaw or Hermes and upstream LLM APIs, mapping virtual keys to real keys stored under /etc/clawshell, scanning request/response bodies for PII (block or redact), and optionally filtering IMAP email by sender allow/deny lists. An onboard wizard rewires one downstream agent to route through the proxy.

How it went

They framed the threat model as OS-level isolation rather than prompt trust: a hijacked agent gets only a scoped virtual identifier with no real credentials or lateral secret access.

Worth knowing

Secrets still live on the same host under Unix permissions; this is not a full VM sandbox. Streaming SSE responses skip DLP scanning.

Save ideas in this browser. Upvotes count across the board.

Prompt excerpt

No prompt excerpt was provided by the source.

Explore all →
docsGrok-BotMeta-Muse

TV De-Ad Bot: Grok Bot strips ads from your own smart TV

Gumroad guide + free Grok Bot marketplace template (also chapters for Meta Muse Code, ChatGPT Dots, OpenClaw, Hermes): reversible adb/Projectivy on Google TV/Android TV/Fire TV, brand-realistic limits on LG/Samsung/Roku, NextDNS/Pi-hole/AdGuard Home, ACR opt-outs, and a weather/calendar screensaver—approve each change, undo included, your device only.

Source ↗