ClawShell

What ClawShell does

ClawShell runs as a privileged process between the agent and providers such as OpenAI, Anthropic and OpenRouter. Real keys sit in /etc/clawshell/clawshell.toml, readable only by the clawshell system user, so the agent never sees them. It maps each virtual key to a provider and injects the right header format for that provider.

A data loss prevention layer scans request and response bodies with regex patterns you define in TOML, each set to block or redact; the README gives SSNs, credit card numbers and emails as examples. Streaming responses pass through without scanning. An email endpoint can expose a mailbox with sender allowlist or denylist rules per virtual key, with Gmail and Outlook presets and manual IMAP. OAuth sign-in for Codex and ChatGPT uses a device code flow with automatic token refresh. A loopback-only /admin/stats endpoint reports protected requests, token counts and emails hidden by the sender policy.

Key features

• Virtual-to-real API key mapping for OpenAI and Anthropic style providers
• PII scanning of requests and responses with custom block or redact regex patterns
• Sender-based email filtering per virtual key, with Gmail and Outlook presets
• OAuth device code login for Codex and ChatGPT with automatic token refresh
• Loopback-only /admin/stats counters persisted to disk

When to use it

• Keeping real API keys out of reach of an agent that runs tools and shell commands
• Blocking or redacting personal data before it reaches a model provider
• Letting an agent read only mail from approved senders

Who it is for: Operators who run Hermes Agent or OpenClaw on a server and want credential isolation and a data loss prevention layer.

How it fits with Hermes Agent

The README names OpenClaw and Hermes Agent together as the ecosystem it protects, and the repository carries hermes and hermes-agent topics.

Note: Streaming (SSE) responses are passed through without PII scanning.

FAQ