Appearance
OpenClaw Integration (GhostClaw)
GhostClaw makes OpenClaw enterprise-safe by ensuring sensitive data never reaches an LLM — without changing how anyone works.
Overview
OpenClaw is a 307K-star open-source personal AI assistant with plugin-based architecture, MCP support, and 3M+ monthly active users. GhostAI integrates with OpenClaw at four levels, from zero-code configuration to deep runtime hooks.
Agent Onboarding
Paste this into any OpenClaw agent to install and configure GhostClaw automatically (replace the URL with your GhostAI instance):
Read https://your-ghostai-instance.com/SKILL.md and follow the instructions to install and configure GhostClaw for OpenClawThe /SKILL.md endpoint is served by every GhostAI backend deployment with URLs pre-configured for that instance.
Context Engine Integration New
GhostClaw integrates with OpenClaw's Context Engine to apply PII tokenization across the entire conversation window — not just the current message. When the Context Engine assembles context from conversation history, memory retrieval, file content, and tool outputs, GhostClaw scans and tokenizes every component before it reaches the LLM.
This matters because PII accumulates across turns. A name in message 1, an address in message 3, a diagnosis in message 7 — all concatenated into a single context payload. Without Context Engine integration, only the current message is protected while prior turns leak through. With GhostClaw's Context Engine wrapper, the full assembled context is tokenized, so the LLM never sees raw PII from any source.
What this means in practice:
- Lower token waste from redundant PII handling per-message
- Consistent tokenization across multi-turn conversations — the same entity gets the same token throughout
- Memory and file retrieval results are PII-scanned before entering the context window
- Agents that use persistent memory (e.g., mem9) store tokenized data, not raw PII
Context Engine integration is available on Business and Enterprise plans.
Integration Tiers
Tier 0: Zero-Code Config (Available Now)
No plugin code needed. Works with any GhostAI plan.
Custom Provider — Add GhostAI as an LLM provider in openclaw.json:
json
{
"models": {
"providers": {
"ghostclaw": {
"api": "openai-completions",
"baseUrl": "https://your-ghostai-instance.com/v1",
"apiKey": "${GHOSTAI_API_KEY}",
"models": ["openai/gpt-4o", "openai/gpt-4o-mini", "anthropic/claude-3.5-sonnet"]
}
}
}
}All chat completions through this provider route through GhostAI with transparent PII tokenization and streaming.
MCP Server — Expose GhostAI tools to the agent:
json
{
"mcp": {
"ghostai": {
"command": "ghostai-mcp",
"env": {
"GHOSTAI_BASE_URL": "https://your-ghostai-instance.com",
"GHOSTAI_API_KEY": "${GHOSTAI_API_KEY}"
}
}
}
}Tools available: ghostai_detect_pii, ghostai_sanitize, ghostai_restore, ghostai_scan_file.
Tier 1: GhostClaw Plugin (Personal+)
Install the full plugin for dynamic model catalog, slash commands, and privacy receipts:
bash
openclaw plugins install ghostclawThen run /ghostclaw setup to configure your API key and industry vertical.
Provider registration with dynamic model catalog from GET /v1/models.
Agent tools:
ghostclaw_scan— Scan text for PII with compliance tagsghostclaw_sanitize_file— Scan files (100+ formats) for PIIghostclaw_compliance_check— Check text against industry compliance rules
Slash commands: /ghostclaw status, /ghostclaw scan, /ghostclaw config, /ghostclaw explain, /ghostclaw setup, and more.
Tier 2: Universal Hooks (Team+)
Hooks intercept every LLM call regardless of provider. The user can select OpenAI, Anthropic, Gemini, or any provider — GhostClaw tokenizes PII before the request reaches that provider and detokenizes in the response.
before_llm_call— Tokenize PII in all user messagesafter_llm_call— Restore original values in LLM response
Context Engine wrapper (Business+) — PII-safe context assembly across the entire conversation window. Every component of the assembled context — history, memory, files, tool outputs — is scanned and tokenized before reaching the LLM. See Context Engine Integration above for details.
Media understanding provider (Business+) — Image, audio, and video PII detection. Scans uploaded media for PII (driver's licenses, credit cards, medical records in images; sensitive content in audio transcriptions) and applies the same tokenization policy as text.
Tier 3: NemoClaw Sandbox (Enterprise)
NemoClaw is NVIDIA's secure agent stack. GhostAI serves as the mandatory inference gateway inside NemoClaw sandboxes:
- All inference routes through GhostAI — direct LLM access is blocked
- Sandbox network policy pre-approves GhostAI endpoints
- PII protection becomes mandatory, not optional
"NemoClaw secures the sandbox. GhostClaw secures the data."
Configuration
| Setting | Description | Default |
|---|---|---|
baseUrl | GhostAI instance URL | http://localhost:8000 |
vertical | Industry vertical | general |
sensitivity | Detection threshold | medium |
feedback | Receipt verbosity (auto/always/verbose/silent) | auto |
failMode | Behavior when GhostAI is unreachable (block/warn/pass) | warn |
Industry Verticals
| Vertical | Compliance Frameworks |
|---|---|
| Healthcare | HIPAA, HITECH, 42 CFR Part 2 |
| Financial | PCI-DSS, SOX, GLBA |
| Legal | Attorney-client privilege, ABA rules |
| Government | CUI, ITAR, FedRAMP, NIST 800-171 |
| Technology | DTSA, trade secrets, SOC 2 |
| Education | FERPA, COPPA |
| Retail | PCI-DSS, CCPA |
| Manufacturing | DTSA, trade secrets, ITAR |
| General | GDPR, CCPA, state privacy laws |
Privacy Receipts
GhostClaw provides three layers of feedback:
- Glance — Single-line badge:
🛡 GhostClaw: 3 items protected · HIPAA · 12ms - Summary — Entity breakdown with types, counts, and confidence
- Detail — Per-entity deep dive with original text, token, detector, and compliance tags
Setup Guide
See the AI Tool Setup Guide for step-by-step instructions.
API Reference
GhostClaw uses the standard GhostAI APIs:
- OpenAI-Compatible API — Chat completions and model listing
- MCP Server — PII detection, sanitization, and file scanning tools
- Detection API — Direct PII detection endpoints