teodorofodocrispin-cmyk/trustboost-pii-sanitizer

1. Repository Role

This is a distribution and integration package for the hosted TrustBoost PII Sanitizer API, rather than the sanitizer service implementation itself. It publishes agent-facing metadata, a browser landing page, and an installable skill directing AI-agent clients to api.trustboost.dev.

Primary user entry points:

• Landing page: index.html

• Agent/plugin descriptors: ai-plugin.json, plugin.json, manifest.json

• Installable skill: skills/trustboost-pii-sanitizer/SKILL.md

2. Hosted Sanitization Boundary

The core processing boundary is external: clients submit text to the TrustBoost API before forwarding it to downstream LLMs. The API returns redacted content, entity metadata, a safety score, risk classification, and quota information.

The package does not contain API handlers, PII detectors, persistence, or Solana integration code. Those belong to the separately linked trustboost-api service. This repository’s integration contract is documented in README.md and conveyed through the skill definition at skills/trustboost-pii-sanitizer/SKILL.md.

3. Agent Skill Integration

The main operational artifact is the TrustBoost skill:

• skills/trustboost-pii-sanitizer/SKILL.md

• SKILL.md

It is intended for AI coding/agent environments installed via ClawHub. The skill instructs an agent to treat sanitization as a pre-LLM inspection stage: identify potentially sensitive outbound text, call the hosted sanitizer, then use only the sanitized result in external model/API requests.

4. Protocol and Discovery Surface

Multiple metadata files expose the same hosted capability to different agent ecosystems:

• ai-plugin.json

• plugin.json

• manifest.json

• molt.json

• glama.json

These descriptors are the repository’s machine-readable integration layer. The externally hosted API supports REST sanitization plus MCP JSON-RPC, Google A2A messaging, ANP discovery, and x402 payment discovery; clients discover or invoke those services through the published metadata rather than local runtime code.

5. Request, Payment, and Proof Flow

The intended control flow is:

1. Agent receives text that may contain PII or credentials.

2. Agent calls hosted preview or paid sanitization endpoints.

3. TrustBoost redacts detected entities and reports policy/risk information.

4. Agent sends sanitized text to its target LLM or external service.

5. Paid requests are authorized through Solana USDC transaction data and generate a verifiable on-chain sanitization proof.

Key external boundaries are the hosted TrustBoost API, Solana mainnet, and Helius for proof anchoring, as described in README.md.

6. Presentation and Package Metadata

index.html is the static public landing surface, likely deployable without an application server. health.json provides a lightweight package/service-status-oriented artifact. Marketplace and package identity are represented by:

• manifest.json

• plugin.json

• ai-plugin.json

There is no evidence in the tree of a backend runtime, database schema, frontend build system, or infrastructure-as-code. benchmark.py is a standalone evaluation artifact, not an apparent production execution path.

7. Policy and Privacy Positioning

The package defines a privacy-oriented agent boundary: raw input should not be retained or propagated after sanitization, while sanitized output and paid-proof metadata support auditing. Supporting policy material is located in:

• PRIVACY.md

• AGENTS.md

• Digital Consciousness Manifesto

For implementation planning, treat the hosted API and its endpoint contract as the system of record; this repository is the client-distribution, discovery, and instruction layer around it.