XSafeClaw

What XSafeClaw does

XSafeClaw treats agent security as a live control problem. Its README argues that agents can be talked into dangerous actions through prompt injection, tool misuse and silent privilege escalation, so it provides monitoring and guarding while agents run instead of log review after the fact. The first release listed Claw Monitor, Safe Chat, Asset Shield, Guard, Agent Office and Onboard Setup.

Hermes support appears in the release notes. Version 1.1.0 added Hermes session-context guard coverage to Runtime Guard, along with approval-request display fixes and budget controls. Version 1.0.8 made per-model token usage and cost accounting uniform across OpenClaw, Hermes and Nanobot, and an April release added discovery of Hermes next to OpenClaw and nanobot, with best-effort startup of installed gateways when the server boots. Releases are published to PyPI, and a view called Agent Valley shows sessions from all three agents together.

Key features

• Runtime Guard with Hermes session-context coverage
• Safe Chat for talking to guarded agents with Markdown rendering
• Per-model token usage and cost accounting across OpenClaw, Hermes and Nanobot
• Budget controls
• Agent Valley view of mixed-runtime sessions
• Discovery of installed Hermes, OpenClaw and nanobot runtimes

When to use it

• Watch what a Hermes agent is doing from a safety dashboard
• Track token cost per model across several agent runtimes
• Apply approval and budget controls to agent sessions

Who it is for: Teams and individuals running agents with real tool access who want runtime monitoring and guarding.

How it fits with Hermes Agent

Hermes Agent is one of three runtimes XSafeClaw supports, next to OpenClaw and Nanobot; it is not built only for Hermes.

FAQ