Skip to content

StevenJohnson998/agent-registry

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

⏸️ PROJECT STATUS: ON HOLD
This project is paused while the maintainer focuses on other priorities.
The architecture and design docs remain available for review and feedback.
Contributions are welcome but may not be reviewed immediately.

Agent Registry

The Trust & Privacy Layer for the Agentic Ecosystem.

License Status ADHP

The Problem

AI agents are multiplying. An orchestrator handling a financial analysis task might have 50 agents to choose from — but no good way to decide:

  • Which agent is the best for this specific task? Not a generic rating — proven quality at this capability.
  • What will it do with my data? Will it train on it? Store it? Forward it to a third party?
  • Is it compliant with the regulations I care about? GDPR, HIPAA, POPIA, CCPA — can I filter by that?
  • Who's behind this agent? Is there a verified, accountable organization operating it?
  • What's the most cost-effective option at the trust level I need?
  • If it delegates to another agent, will the privacy and compliance chain hold?

Today, registries catalog agents. None of them answer these questions. Agent Registry is an open-source trust layer that does. We don't replace MCP or A2A — we complement them. MCP connects agents to tools. A2A lets agents talk to each other. We help agents decide who to trust.


What We're Building

1. Agent Data Handling Policy (ADHP)

A standardized way for agents to declare what happens to your data — and for orchestrators to filter agents based on those declarations.

Five levels, from open (no guarantees) to zero-trace (streaming only, no storage, no logs, no delegation):

Level Training Persistence Logging Delegation Third-Party Sharing
open Yes Unlimited Full Unrestricted Allowed
standard No Session Metadata only Same level+ With consent
sensitive No Request only Metadata only Declared only Anonymized only
strict No Request only None Same level+ Not allowed
zero-trace No None None None Not allowed

Key innovation: Delegation Cascading. When Agent A delegates to Agent B, the data handling level must be maintained or strengthened — never weakened. The registry validates the entire chain.

Privacy legislation compliance tags. Agents declare which regulations they comply with — GDPR, HIPAA, CCPA, POPIA, PIPEDA, LGPD, and others. Orchestrators filter by regulation: "find me an agent compliant with GDPR and HIPAA that handles data at strict level."

📄 Full ADHP Specification

2. Operator Verification (KYC)

Trust requires accountability. We verify the organizations behind agents through tiered verification:

Tier What's Verified Badge
Unverified Nothing — self-registered
Email-verified Valid email confirmed ✉️
Identity-verified Legal identity confirmed via documents
Audited Identity + technical ADHP compliance audit 🛡️

Orchestrators can combine filters: "find me an agent with ADHP ≥ sensitive AND verification ≥ identity-verified."

📄 Full KYC Specification

3. Per-Capability Trust Scoring

A flat rating is useless. An agent might be excellent at financial analysis but poor at creative writing. Our registry stores contextual trust maps — quality and confidence scores per capability.

GET /agents?capability=financial-analysis&min_quality=4.0&adhp_level=strict&jurisdiction=EU&compliance=GDPR

How It Fits In the Ecosystem

We're not competing with existing projects — we're filling a gap that none of them address:

┌─────────────────────────────────────────────────────┐
│                   Agentic Ecosystem                  │
│                                                      │
│  MCP Registry ─── catalogs servers/tools             │
│  A2A Protocol ─── agents discover & talk to agents   │
│  Kong/Solo.io ─── enterprise gateway & governance    │
│                                                      │
│  Agent Registry ─── WHO DO I TRUST? ◄── you are here │
│    ├── ADHP: what happens to my data?                │
│    ├── Compliance: GDPR, HIPAA, CCPA, POPIA...      │
│    ├── KYC: who's behind this agent?                 │
│    ├── Trust scores: is it good at THIS task?        │
│    └── Delegation chains: will trust hold?           │
└─────────────────────────────────────────────────────┘

We plan to federate with the official MCP Registry — pulling server metadata and adding our trust layer on top. We support both MCP and A2A protocols. We're compatible with AAIF standards.


Project Status

Alpha MVP — working registry with REST API, PostgreSQL storage, API key auth, rate limiting, search with filtering, web dashboard, and Docker deployment. 47 tests, 84% coverage.

Component Status
REST API (FastAPI) ✅ Working
Agent CRUD + search ✅ Working
ADHP fields + filtering ✅ Working
API key auth (writes) ✅ Working
Rate limiting (Redis) ✅ Working
Web dashboard ✅ Working
Docker deployment ✅ Working
ADHP Specification 📄 v0.1 Draft
KYC Specification 📄 v0.1 Draft
Operator system 🔨 Next
Public deployment 🔨 Next

See the Roadmap for what's coming.


Quick Start

# Clone
git clone https://github.com/StevenJohnson998/agent-registry.git
cd agent-registry

# Start (requires Docker)
./batch/run-dev.sh

# API available at http://localhost:8000
# Dashboard at http://localhost:8000/dashboard
# Docs at http://localhost:8000/docs

# Run tests
./batch/run-tests.sh

# Seed example agents
./batch/seed-db.sh

Register an Agent

curl -X POST http://localhost:8000/agents \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "FinanceAnalyzer Pro",
    "description": "Financial document analysis with strict privacy",
    "capabilities": ["financial-analysis", "document-parsing"],
    "endpoint": "https://api.finanalyzer.example.com",
    "protocol": "REST",
    "data_handling": {
      "level": "strict",
      "training_opt_out": true,
      "max_retention": "request",
      "content_logging": false,
      "delegation_policy": "same_or_higher",
      "jurisdiction": ["DE", "FR"],
      "compliance": ["GDPR"],
      "third_party_sharing": { "enabled": false }
    }
  }'

Find Agents by Trust Requirements

# Find strict+ agents in the EU, GDPR-compliant, that don't share data
curl "http://localhost:8000/agents?min_data_handling_level=strict&jurisdiction=EU&compliance=GDPR&third_party_sharing=false"

Repository Structure

agent-registry/
├── docs/                    # Specifications
│   ├── ADHP-SPEC.md         # Agent Data Handling Policy specification
│   ├── KYC-SPEC.md          # Operator verification specification
│   └── ROADMAP.md           # Project roadmap
├── dev/                     # Application source code
│   ├── app/                 # FastAPI application
│   │   ├── api/             # API routes
│   │   ├── core/            # Config, auth, rate limiting
│   │   ├── models/          # SQLAlchemy models
│   │   └── schemas/         # Pydantic schemas
│   ├── tests/               # Test suite
│   ├── alembic/             # Database migrations
│   ├── static/              # Web dashboard
│   └── docker-compose.dev.yml
├── batch/                   # Operational scripts
│   ├── run-dev.sh
│   ├── stop-dev.sh
│   ├── run-tests.sh
│   ├── seed-db.sh
│   └── ...
└── notes/                   # Internal project notes

Contributing

We're especially looking for feedback on:

  • ADHP Specification — Are the five levels right? What data handling concerns are we missing?
  • Ecosystem integration — How should we best work with MCP and A2A?
  • Real-world testing — Do you operate agents? We'd love to register them and test our assumptions.

How to Contribute

  1. Read the ADHP Spec and KYC Spec
  2. Open an Issue or Discussion with your feedback
  3. Submit PRs for code or spec improvements

All contributions are welcome, from spec feedback to code to documentation.


License

Apache 2.0 — see LICENSE.


Links

About

An MCP-first service discovery registry for autonomous AI agents.

Topics

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages