Skip to content

GhostClaw Demo Script ​

Seven demos. Fifteen minutes. Zero PII reaches the LLM.

This script is designed for live customer demos, conference booth presentations, and internal stakeholder walkthroughs. Each demo is self-contained with exact inputs, expected outputs, and talking points.


Pre-Demo Checklist ​

Run through this before every demo:

  • [ ] GhostAI instance is running and reachable
  • [ ] API key is valid: curl -s -H "Authorization: Bearer $GHOSTAI_API_KEY" https://your-ghostai-instance.com/v1/models | head -1
  • [ ] OpenClaw is installed with GhostClaw plugin active: openclaw plugins list
  • [ ] GhostClaw is configured: /ghostclaw config shows valid settings
  • [ ] Feedback mode set to verbose: /ghostclaw config feedback verbose
  • [ ] Vertical set to general: /ghostclaw config vertical general
  • [ ] Test file demo-patient-record.txt is ready (see below)
  • [ ] Test image demo-id-card.png is ready (use sample from GhostAI docs)
  • [ ] Internet connection is stable (or use local GhostAI instance)
  • [ ] Terminal font size is large enough for audience visibility

Create test data ​

bash
cat > demo-patient-record.txt << 'EOF'
PATIENT INTAKE FORM
====================
Name: Maria Garcia
Date of Birth: 04/22/1988
Social Security: 456-78-9012
Medical Record Number: MRN-2026-44891
Phone: (312) 555-0147
Email: maria.garcia@email.com
Insurance: BlueCross ID BC-445-991-20
Primary Diagnosis: Type 2 Diabetes Mellitus (E11.9)
Medications: Metformin 500mg BID, Lisinopril 10mg daily
Emergency Contact: Carlos Garcia, (312) 555-0148
EOF

Demo 1: The Accidental Leak ​

Duration: 2 minutes Story: A user pastes a customer record into the AI, not thinking about the sensitive data it contains.

Input ​

Type (or paste) this into OpenClaw:

Hey, can you help me draft a follow-up email to this customer?

Name: Sarah Johnson
Email: sarah.johnson@acmecorp.com
Phone: (555) 867-5309
SSN: 234-56-7890
Account: 4532-1234-5678-9012 exp 03/28
Address: 742 Evergreen Terrace, Springfield, IL 62704

Expected Output ​

The AI responds with a natural follow-up email draft, referencing "Sarah" by name. Below the response:

🛡 GhostClaw Privacy Receipt
  ├─ 👤 person           ×1  (97%)  → tokenized
  ├─ ✉️  email address    ×1  (98%)  → tokenized
  ├─ 📞 phone number     ×1  (95%)  → tokenized
  ├─ 🔐 SSN              ×1  (99%)  → tokenized
  ├─ 💳 credit card      ×1  (97%)  → tokenized
  ├─ 📅 expiration date  ×1  (92%)  → tokenized
  └─ 📍 address          ×1  (94%)  → tokenized
  Compliance: PCI-DSS Req 3, CCPA §1798.140
  Scan time: 18ms

Talking Points ​

"Notice what just happened. The user pasted a customer record with an SSN, credit card, and home address. In any other setup, all of that just went straight to OpenAI's servers. With GhostClaw, the LLM never saw any of it. It saw tokens like [SSN_r8t5] and [CC_n6j3]. The response still makes sense because GhostClaw restores the tokens on the way back."

"This is invisible protection. The user didn't do anything special. They didn't click a button or run a command. GhostClaw caught 7 PII entities in 18 milliseconds."


Demo 2: See What Was Protected ​

Duration: 1 minute Story: Show the audience the three layers of privacy feedback.

Steps ​

Layer 1 — Glance badge (already visible from Demo 1):

🛡 GhostClaw: 7 items protected · PCI-DSS · 18ms

Layer 2 — Summary:

/ghostclaw status

Shows the entity breakdown with types, counts, confidence, and compliance tags.

Layer 3 — Deep dive:

/ghostclaw explain

Shows per-entity details: original text, replacement token, detector source, confidence factors, and relevant regulation citations.

Talking Points ​

"GhostClaw uses progressive disclosure. Most users just see the green shield badge — that's enough to know they're protected. IT admins and compliance teams can drill deeper. The /ghostclaw explain view shows exactly what was caught, what detector found it, and which regulations apply. This is the kind of audit trail compliance teams need."


Demo 3: Scan Before You Share ​

Duration: 2 minutes Story: A user wants to check a document for PII before uploading it to the AI.

Steps ​

Reference the test file created in the checklist:

Can you scan demo-patient-record.txt for sensitive data before I share it?

Expected Output ​

The agent calls ghostai_scan_file and returns:

Scan Results: demo-patient-record.txt
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Found 9 PII entities:

  👤 PERSON          "Maria Garcia"           line 3   (97%)
  📅 DATE_OF_BIRTH   "04/22/1988"             line 4   (94%)
  🔐 US_SSN          "456-78-9012"            line 5   (99%)
  ⚕️ MEDICAL_RECORD   "MRN-2026-44891"         line 6   (95%)
  📞 PHONE_NUMBER    "(312) 555-0147"         line 7   (96%)
  ✉️ EMAIL_ADDRESS    "maria.garcia@email.com" line 8   (98%)
  🏥 INSURANCE_ID    "BC-445-991-20"          line 9   (91%)
  ⚕️ DIAGNOSIS_CODE   "E11.9"                  line 10  (88%)
  👤 PERSON          "Carlos Garcia"          line 12  (96%)

Compliance: HIPAA §164.514, CCPA §1798.140
Risk Level: HIGH — contains PHI

Talking Points ​

"Before the file even gets near the AI, we know exactly what's in it. Nine PII entities, including PHI — protected health information. GhostClaw flagged the diagnosis code, the medical record number, and the insurance ID. These are the kinds of entities that generic PII tools miss entirely."

"If this were a healthcare organization, the HIPAA compliance tag tells the compliance officer exactly which regulation applies."


Demo 4: Image Intelligence ​

Duration: 2 minutes Story: A user uploads a photo of an ID card without thinking about the PII it contains.

Steps ​

Can you read the information from this ID card?

Attach demo-id-card.png (a sample driver's license image from the GhostAI docs).

Expected Output ​

GhostClaw's media understanding provider intercepts the image:

🛡 GhostClaw: Image PII detected before LLM processing

  🪪 US_DRIVER_LICENSE  detected (94% confidence)
     State: California
     Contains: name, DOB, address, license number, photo

  Actions taken:
  ├─ Text fields: tokenized (name, DOB, address, license #)
  ├─ Face: de-identified (blur applied)
  ├─ EXIF GPS: stripped
  └─ Barcode: decoded and scanned

  The LLM received a redacted version of this image.

Talking Points ​

"This is where GhostClaw goes far beyond text scanning. The user uploaded a photo of a driver's license. GhostClaw recognized it as a California DL — it knows templates for all 50 US states plus passports. It tokenized the text fields, blurred the face, stripped GPS data from the image metadata, and decoded the barcode on the back. The LLM got a sanitized version."

"No other OpenClaw privacy plugin can do this. Most can't even scan text reliably."


Demo 5: Compliance Tags ​

Duration: 1 minute Story: Switch to a healthcare vertical to show regulation-specific tagging.

Steps ​

/ghostclaw config vertical healthcare

Then send:

Patient John Doe, MRN 78432, diagnosed with Major Depressive Disorder on 02/15/2026. 
Prescribed Sertraline 50mg daily. Insurance: Aetna HMO, member ID AET-992-441.

Expected Output ​

🛡 GhostClaw Privacy Receipt
  ├─ 👤 PERSON           ×1  "John Doe"        → tokenized
  ├─ ⚕️ MEDICAL_RECORD    ×1  "78432"           → tokenized
  ├─ ⚕️ DIAGNOSIS         ×1  "Major Depre..."  → tokenized
  ├─ 📅 DATE             ×1  "02/15/2026"      → tokenized
  ├─ 💊 MEDICATION        ×1  "Sertraline 50mg" → tokenized
  └─ 🏥 INSURANCE_ID     ×1  "AET-992-441"     → tokenized
  Compliance: HIPAA §164.514(a), HIPAA §164.514(b), 42 CFR Part 2
  Scan time: 21ms

Talking Points ​

"Notice the compliance tags changed. In healthcare vertical, GhostClaw applies HIPAA-specific rules. It caught the diagnosis — Major Depressive Disorder falls under 42 CFR Part 2 for substance abuse and mental health records, which has even stricter protections than standard HIPAA. The medication name is flagged because it reveals the diagnosis."

"This is the kind of vertical intelligence that matters for regulated industries. A generic PII scanner sees a name and a date. GhostClaw sees PHI with specific regulatory implications."


Demo 6: Works With Any Provider ​

Duration: 1 minute Story: Switch LLM providers to prove that GhostClaw protects all of them.

Steps ​

If using the plugin with hooks (Tier 2), simply switch providers in OpenClaw's model picker:

Switch from OpenAI GPT-4o → Anthropic Claude 3.5 Sonnet

Send the same test message:

My name is Alex Rivera, SSN 111-22-3333, and I need help with my tax return.

Expected Output ​

Same privacy receipt, same tokenization, different provider:

🛡 GhostClaw: 2 items protected · CCPA · 15ms

Talking Points ​

"This is critical. GhostClaw isn't tied to one provider. Whether you're using OpenAI, Anthropic, Google, Mistral, or a local Ollama model — the protection is the same. The before_llm_call hook intercepts every request regardless of where it's going."

"For enterprises, this means one privacy layer covers their entire AI stack. They don't need separate solutions for each provider."


Demo 7: The NemoClaw Story ​

Duration: 2 minutes Story: Show how NemoClaw's sandbox makes GhostClaw protection mandatory.

Steps ​

If running in a NemoClaw sandbox, show the network policy:

bash
nemoclaw status

Expected:

Sandbox: demo-agent
Status: running
Network policy: restricted
  ✓ ghostai-instance.com:443 (allowed)
  ✗ api.openai.com (blocked)
  ✗ api.anthropic.com (blocked)
  ✗ * (default deny)
Privacy: GhostClaw enforced via network policy

Then demonstrate a bypass attempt — try to use a direct OpenAI provider:

/model openai/gpt-4o

The agent tries to reach api.openai.com directly:

⚠ Connection blocked by sandbox policy.
  Agent attempted: api.openai.com:443
  Reason: Not in allowed egress list
  Resolution: All inference must route through GhostAI

Talking Points ​

"This is the complete enterprise story. NemoClaw's sandbox has a strict network policy — the only allowed outbound connection is to GhostAI. The agent literally cannot reach OpenAI, Anthropic, or any other LLM directly. Every request must go through the privacy proxy."

"This isn't optional configuration that a user can turn off. It's enforced at the network level. Even if someone tries to bypass GhostClaw, the sandbox blocks it. This is what enterprise security teams want to hear: mandatory, enforceable, auditable privacy."

"NemoClaw secures the sandbox. GhostClaw secures the data. Together, you have a defense-in-depth stack that passes security review."


Recovery Steps ​

Things go wrong in live demos. Here's how to recover:

GhostAI is unreachable ​

/ghostclaw config failMode warn

The demo continues with a warning banner. Explain that failMode controls degradation behavior.

Plugin isn't responding ​

Restart OpenClaw:

bash
openclaw --reset-plugins

Or disable and re-enable:

bash
openclaw plugins disable ghostclaw && openclaw plugins enable ghostclaw

PII not detected ​

Check settings:

/ghostclaw config

Reset to demo defaults:

/ghostclaw config vertical general
/ghostclaw config sensitivity high
/ghostclaw config feedback verbose

NemoClaw sandbox won't start ​

bash
docker restart $(docker ps -q --filter name=nemoclaw)
nemoclaw start

Wrong model selected ​

/model ghostclaw/openai/gpt-4o

Audience asks about latency ​

Show real-time stats:

/ghostclaw status

Point to the "Scan time" value. Typical: 10–30ms. Always under 50ms p95.


Demo Flow Cheat Sheet ​

#DemoDurationKey Moment
1The Accidental Leak2 min7 PII entities caught silently
2See What Was Protected1 min3-layer progressive disclosure
3Scan Before You Share2 minFile scanning with 9 entities
4Image Intelligence2 minID card detection + face blur
5Compliance Tags1 minHIPAA tagging on healthcare data
6Any Provider1 minSame protection across providers
7NemoClaw Sandbox2 minMandatory enforcement via policy
Total~11 min

Frequently Asked Questions (From Demos) ​

"Does this slow things down?"

Scan latency is under 50ms p95. For perspective, the LLM response takes 2–10 seconds. You won't notice 20ms.

"What if GhostAI is down?"

failMode controls this. In block mode, messages are held. In warn mode, they go through with a warning. Enterprises typically use block for compliance.

"Can users turn this off?"

With Tier 0/1, the user chooses the provider. With Tier 2 hooks, it's on for all providers. With NemoClaw (Tier 3), it's enforced by network policy — users can't bypass it.

"What about on-prem?"

GhostAI can be self-hosted. Everything runs in your infrastructure. No data leaves your network.

"How is this different from OpenClaw's built-in privacy?"

OpenClaw's native privacy.redactPII handles phone numbers and user ID hashing. GhostClaw handles 64+ entity types, files, images, audio, video, compliance tagging, DSAR, and reversible tokenization. They're in different leagues.

"What does it cost?"

The GhostClaw plugin is free and open source. GhostAI pricing: Free (self-hosted), Personal ($19/mo), Team ($49/user/mo), Enterprise (custom).


For installation from scratch, see the Full Lab Guide. For advanced configuration, see the Configuration Guide.

SOC 2 Type II Certified | BAA Available | HIPAA-Ready