Vault Form Agent

Vault Form Agent is an enterprise-grade prototype for secure, approval-gated document-to-form automation. It combines a metadata-rich document locker, scalable PHI/PII redaction policies, a Claude-style chat agent, official-site research, browser form execution, and audit logging.

The system is designed around HIPAA/PHI/PII-aware handling patterns. It is not a certified compliance product by itself; production HIPAA compliance also requires hosting controls, BAAs, policies, training, risk assessments, and operational governance.

Highlights

• Chat-first agent workspace with document scope controls.
• Document locker for any file type with name, description, tags, metadata, sensitivity, and consent.
• Docling-ready extraction path with OCR fallbacks for PDFs and scanned images.
• Built-in and custom redaction rules for PHI, PII, financial, government, and internal identifiers.
• Official website research agent for prompts without known target URLs.
• Approval review before execution.
• Visible browser automation for typing, clicking, selecting, uploading, and pausing before final submit.
• Audit trail for document access, field use, approvals, and execution.
• CLI and agent-skill scaffold for Codex/Claude-style integrations.

Demo

Primary end-to-end recording:

• demo-recordings/system-end-to-end-google-form-fill.mp4
• Corrected clean browser-fill proof: demo-recordings/google-form-clean-8of8-submit-guard.webm
• Proof screenshot: demo-recordings/google-form-clean-8of8-submit-guard.png

The recording demonstrates:

1. Agent chat workflow.
2. Real-world Google Form test case.
3. Approval-gated execution.
4. Backend-launched Chrome session.
5. Real fields filled on a public Google Form.
6. Submit guard blocks final submission and leaves review to the user.

Sample files are in demo-data/.

Visual Walkthrough

Agent chat workspace

The user selects document scope, enters a natural-language request, reviews the agent plan, and prepares the browser execution from a Claude-style chat surface.

Real Google Form filled by the agent

The browser runner opens a public Google Form and prepares the visible fields using the planned values.

Eight-field controlled browser proof

The corrected proof run shows eight fields prepared and final submit blocked by the safety guard.

Full workflow recording frames

These frames show the broader locker-to-agent-to-browser flow used in the demo recording.

Architecture

flowchart TD
    UI[React Enterprise Web App] --> API[FastAPI Backend]
    API --> Locker[Document Locker]
    API --> Redaction[Redaction Policy Engine]
    API --> Chat[Chat Agent Orchestrator]
    API --> Research[Official Site Research Agent]
    API --> Approval[Approval Workflow]
    API --> Audit[Audit Trail]
    Approval --> Browser[Visible Browser Agent]
    Locker --> Extract[Docling/OCR Extraction]
    Locker --> Store[(Document/Object Store Target)]
    Redaction --> Vault[(Secure Token Vault Target)]
    Chat --> Scope[Document Scope Enforcement]

Local Run

Backend:

cd backend
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --host 127.0.0.1 --port 8010

Frontend:

cd frontend
npm install
npm run dev -- --port 5173

Open:

• App: http://localhost:5173
• Backend health: http://127.0.0.1:8010/health

Docker compose is included for the target service topology. The current desktop demo uses local backend port 8010 so it does not collide with other services on 8000.

Product Flow

1. Upload documents in Locker.
2. Add document name, description, tags, metadata, sensitivity, and consent.
3. Configure redaction rules if custom fields need masking.
4. Open Agent Chat.
5. Choose document scope:
   • no documents
   • selected documents
   • all non-sensitive documents
   • all documents including sensitive
6. Enter or speak a prompt.
7. Agent researches official sites when needed.
8. Agent drafts field mappings and asks for review.
9. User approves.
10. Backend opens Chrome, fills the form, and stops before final submit.

Repository Layout

backend/            FastAPI service, agents, redaction engine, browser runner
frontend/           React + TypeScript enterprise web app
cli/                Command-line API client
demo-data/          Synthetic sample documents
demo-recordings/    Demo videos and screenshots
docs/               Architecture, handbook, compliance, and demo guides
skills/             Agent skill scaffold
legacy_streamlit/   Original MVP preserved for reference

Documentation

• Architecture
• Security and Compliance Notes
• Redaction Handbook
• Browser Agent Handbook
• Demo Guide
• API and Agent Integration
• GitHub Publishing Notes

API Overview

• GET /api/locker/documents
• POST /api/locker/documents
• POST /api/redaction/rules
• GET /api/redaction/rules
• POST /api/redaction/preview
• POST /api/agent/drafts
• POST /api/agent/research
• POST /api/agent/approvals/{approval_id}/approve
• POST /api/agent/approvals/{approval_id}/execute
• GET /api/audit/events
• GET /api/security/locker-policy

Safety Boundaries

The browser agent does not click final submit automatically. It pauses for:

• CAPTCHA
• MFA
• payment
• signature
• legal attestation
• external file upload confirmation
• final submit

Validation Used

• Backend compile: python -m compileall backend/app
• Frontend build: npm run build
• API smoke tests for redaction, upload, scoped draft, approval, and browser execution
• Browser UI smoke tests for Locker, Agent Chat, Redaction, and Approval

License

Add a license before public distribution.