Agent Trust

GARL Protocol

Prove what your AI agent was authorized to do — and what it actually did.

Capability tokens set hard limits on an agent — spend caps, merchant allowlists, side-effect class, expiry — and a delegated token can only narrow its parent, never widen it.
Every action becomes an ECDSA-secp256k1-signed Action Receipt bound to the token that authorized it, Merkle-anchored on Base mainnet, and verifiable offline without trusting GARL.

Add your agent · Anchor chain · Website · Docs · Live receipt · Try It

---

---

Try it now

Path A — For Agents (SDK / MCP)

With Claude Desktop or Cursor (MCP)

Add to your Claude Desktop config (claude_desktop_config.json) or Cursor MCP settings:

{
  "mcpServers": {
    "garl": {
      "command": "npx",
      "args": ["-y", "@garl-protocol/mcp-server"]
    }
  }
}

That's it — 29 named tools (including batch variants like garl_verify_batch) are now available in your AI assistant: receipts, Trust Vector lookups, capability tokens (issue/verify/revoke), Capability Gate pre-flight, UETA §10(b) undo, and more.

With curl (zero install)

# Check an agent's trust score
curl -s "https://api.garl.ai/api/v1/trust/verify?agent_id=5872ce17-5718-4980-ade3-e51c9556fb53" | python3 -m json.tool

# Find the most trusted coding agent
curl -s "https://api.garl.ai/api/v1/trust/route?category=coding&min_tier=silver" | python3 -m json.tool

# See the live leaderboard
curl -s "https://api.garl.ai/api/v1/leaderboard?limit=5" | python3 -m json.tool

With Python

pip install garl-protocol

import garl

garl.init("your_api_key", "your_agent_uuid")
garl.log_action("Analyzed dataset", "success", category="data")

result = garl.is_trusted("target_agent_uuid", min_score=60)
if result["trusted"]:
    print(f"Safe to delegate — score: {result['score']}/100")

With JavaScript

npm install @garl-protocol/sdk

import { init, logAction, isTrusted } from "@garl-protocol/sdk";

init("your_api_key", "your_agent_uuid", "https://api.garl.ai/api/v1");
await logAction("Generated REST API", "success", { category: "coding" });

const result = await isTrusted("target_agent_uuid", { minScore: 60 });
if (result.trusted) {
  console.log(`Safe to delegate — score: ${result.score}/100`);
}

Capability tokens — authorization with hard limits

# Issue a scoped token for your agent (owner API key required)
curl -s -X POST https://api.garl.ai/api/v1/capability/issue \
  -H "x-api-key: $GARL_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "agent_id": "your-agent-uuid",
    "scope": "payment:stripe.com",
    "side_effect_class": "reversible",
    "spend_limit_usd": 50,
    "merchant_allowlist": ["stripe.com"],
    "expires_in_seconds": 3600
  }' | python3 -m json.tool

# Anyone can verify a token — no auth, no account
curl -s -X POST https://api.garl.ai/api/v1/capability/verify \
  -H "Content-Type: application/json" \
  -d '{"token": "<the JWT-form token>"}' | python3 -m json.tool

A delegated child token can only narrow its parent (lower spend limit, subset allowlist, equal-or-narrower scope, same-or-earlier expiry) — enforced at issue time and re-checked link-by-link at verification. Full wire format: protocol/spec/capability-token-v0.1.md.

Path B — For Code (GitHub Action, 5 lines of YAML)

Sign every AI-authored commit in your pull requests.

# .github/workflows/garl-receipt.yml
name: GARL Receipt
on:
  pull_request:
    types: [opened, synchronize, reopened]
jobs:
  sign:
    runs-on: ubuntu-latest
    permissions: { contents: read, pull-requests: write, checks: write }
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: Garl-Protocol/garl-receipt-action@v1.1.0
        with:
          garl-api-key: ${{ secrets.GARL_API_KEY }}
          garl-agent-id: ${{ secrets.GARL_AGENT_ID }}

Every PR gets a rolling GARL Receipt comment + informational check:

🔐 GARL Verified AI Code
├── Model: claude-opus-4-6
├── Tool: Claude Code
├── Files touched: 12
├── Duration: 4m 12s
├── Signed: ECDSA-secp256k1 ✓
└── Receipt: https://garl.ai/r/a8f3c2d1

Setup guide: Garl-Protocol/garl-receipt-action · Live landing page: garl.ai/for-code.

---

Receipts — a paste-ready proof for every trace

Every submitted trace gets a public shareable Receipt URL at https://garl.ai/r/{short} — a cryptographic proof card (agent, tier, task, duration, SHA-256 hash, ECDSA signature) with an Open Graph image that previews richly in Slack, Twitter/X, GitHub PRs, and LinkedIn.

curl -s https://api.garl.ai/api/v1/verify/6ff83db8 | python3 -m json.tool
#  → receipt_url: https://garl.ai/r/6ff83db8

SDKs expose receipt_url / receiptUrl on every log_action / verify return and a client.receipt(hash) shortcut. The MCP tool garl_receipt resolves any short or full hash to a paste-ready URL.

GitHub Action — sign every AI-authored commit

Add Garl-Protocol/garl/integrations/github-action-receipt to your PR workflow. It detects Claude Code, Cursor, GitHub Copilot, Aider, and Codex co-author trailers, submits a signed trace per qualifying commit, and posts a rolling PR comment + informational check with receipt URLs:

- uses: Garl-Protocol/garl/integrations/github-action-receipt@main
  with:
    garl-api-key: ${{ secrets.GARL_API_KEY }}
    garl-agent-id: ${{ secrets.GARL_AGENT_ID }}

Full setup in integrations/github-action-receipt. Only metadata is uploaded — never diffs or source.

Why GARL?

Problem	GARL's Answer
"What was this agent allowed to do?"	Capability tokens: spend_limit_usd, merchant_allowlist, side_effect_class, expiry — with Biscuit-style attenuation (delegation can only narrow, re-checked link-by-link at verify)
"Did it stay inside those limits?"	Every Action Receipt binds capability_request.token_hash + policy_decision into the signed envelope; the Capability Gate escalates low-trust irreversible actions to a human
"Is this agent reliable?"	5-dimensional trust scoring with Exponential Moving Average
"Which agent should I pick?"	Smart routing by category + minimum certification tier
"Can I verify its track record?"	Immutable ledger with ECDSA-signed execution traces + shareable Receipt URLs
"Does it work with my stack?"	MCP Server · A2A Protocol · REST API · Python & JS SDKs · GitHub Action
"Prove this AI commit is real"	GitHub Action posts a signed receipt per AI-authored commit
"What about on-chain agents?"	ERC-8004 format compatible (off-chain). Receipt-batch Merkle roots are anchored on Base mainnet (MerkleAnchor at 0xBeD7EdeFbEb02be9682bCdeC5fb5D7DA28b1b6F2).

---

Works with

Claude Desktop · Cursor · Any MCP Client · Google A2A · ERC-8004 · REST API · Python · JavaScript · LangChain · CrewAI · AutoGen · LlamaIndex · Semantic Kernel · GitHub Actions

---

How it works

Every agent action is hashed, signed, scored across five dimensions, and made queryable — creating a verifiable trust record.

Agent executes task → SHA-256 hash + ECDSA signature → 5D EMA scoring → Tier assigned → Queryable via API/MCP/A2A

┌─────────────────────────────────────────────────────────────────┐
│                        GARL Protocol                            │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐    │
│  │  Python   │   │   JS     │   │   MCP    │   │   A2A    │    │
│  │   SDK     │   │   SDK    │   │  Server  │   │ JSON-RPC │    │
│  └────┬─────┘   └────┬─────┘   └────┬─────┘   └────┬─────┘    │
│       │              │              │              │            │
│       └──────────────┴──────────────┴──────────────┘            │
│                          │                                      │
│                    ┌─────▼─────┐                                │
│                    │  FastAPI  │  REST + A2A + MCP              │
│                    │  Backend  │  Rate Limited + CORS            │
│                    └─────┬─────┘                                │
│                          │                                      │
│          ┌───────────────┼───────────────┐                      │
│          │               │               │                      │
│    ┌─────▼─────┐  ┌─────▼─────┐  ┌─────▼─────┐               │
│    │ Reputation│  │  Signing  │  │  Webhook  │               │
│    │  Engine   │  │  Engine   │  │  Engine   │               │
│    │ • 5D EMA  │  │ • SHA-256 │  │ • HMAC    │               │
│    │ • Tiers   │  │ • ECDSA   │  │ • Retry   │               │
│    └───────────┘  └───────────┘  └───────────┘               │
│                          │                                      │
│                    ┌─────▼─────┐                                │
│                    │ Supabase  │  PostgreSQL + RLS              │
│                    │           │  Immutable Triggers            │
│                    └───────────┘                                │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

---

ERC-8004 Compatibility

GARL Protocol serves agent metadata in ERC-8004 format (off-chain). Separately, the Merkle roots of batched Action Receipts are anchored on Base mainnet (MerkleAnchor contract 0xBeD7EdeFbEb02be9682bCdeC5fb5D7DA28b1b6F2, chain 8453). Individual receipts are not written on-chain; anyone can verify a receipt's inclusion against an anchored root via verifyProof.

# Get ERC-8004 compatible metadata for any agent
curl -s "https://api.garl.ai/api/v1/agents/{agent_id}/erc8004" | python3 -m json.tool

# Get trust scores in ERC-8004 Reputation Registry feedback format
curl -s "https://api.garl.ai/api/v1/agents/{agent_id}/erc8004/feedback" | python3 -m json.tool

GARL uses the same cryptographic curve as Ethereum (ECDSA-secp256k1), making trust attestations natively verifiable by on-chain systems.

---

Documentation

Topic	Link
Capability Token wire format (spec)	protocol/spec/capability-token-v0.1.md
Action Receipt wire format (spec)	protocol/spec/action-receipt-v0.1.md
Anchoring runbook (weekly Merkle anchor on Base)	docs/runbooks/anchoring.md
Full API Reference (60+ REST endpoints + A2A + MCP)	docs/api-reference.md
MCP Server (29 named tools, including batch variants)	garl.ai/docs#mcp-server
A2A Protocol Integration	garl.ai/docs#a2a
ERC-8004 Compatibility	garl.ai/docs#erc-8004
Python & JS SDKs	garl.ai/docs#sdks
Architecture & Tech Stack	docs/architecture.md
Deployment & Self-hosting	docs/deployment.md
Security	docs/security.md

Interactive API explorer: api.garl.ai/docs (Swagger) · api.garl.ai/redoc

---

Live now

• garl.ai — Live dashboard & real-time trust feed
• Add your agent — Connect any agent (REST, SDK, MCP, GitHub Action) in three steps
• My Agents — sign in (Clerk) and claim the agents you've connected by API key to track their activity from one place
• Registry — Browse connected agents and their signed, verifiable activity
• Verify — Public cryptographic trace verification
• Playground — Interactive API explorer
• Simulator — 5D trust score calculator with what-if analysis
• Compare — Side-by-side agent comparison with radar overlay
• Swagger — Full OpenAPI documentation
• Anchors — every Merkle batch with its root, receipt count, and Base tx (GET /api/v1/anchors)
• MerkleAnchor on Base — Receipt-batch Merkle roots anchored on Base mainnet (chain 8453)
• MCP Registry — Listed as io.github.Garl-Protocol/agent-trust

---

Contributing

GARL Protocol is open source under the Apache 2.0 License. Contributions are welcome — see CONTRIBUTING.md for guidelines and CODE_OF_CONDUCT.md for community standards. Every commit must be DCO-signed (git commit -s).

Requirements: Python 3.10+ for the backend (PEP 604 union syntax), Node 18+ for the frontend. macOS users: the system python3 is 3.9 and will fail backend tests — install 3.10+ via pyenv / brew install python@3.12 and invoke explicitly (python3.12 -m pytest tests/).

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Run tests (python3.12 -m pytest for backend, npx next build for frontend)
4. Commit your changes with DCO sign-off (git commit -s -m 'Add amazing feature')
5. Open a Pull Request

---

Canonical registry, self-hosting, and marks

• Canonical registry: https://api.garl.ai — the single deployment whose public key anchors the GARL Verified status. Public keys are published at /.well-known/garl-keys.json.
• Self-hosting is supported and documented in docs/self-host.md. Self-hosted deployments are first-class participants but are not the canonical registry; see GOVERNANCE.md.
• Trademark policy: TRADEMARK.md. The source code is Apache 2.0; the GARL name and logo are project marks and subject to the policy.

Project decision-making, breaking-change process, and the boundary between repository features (Apache 2.0 forever) and potential future Cloud-only services on the canonical registry are documented in GOVERNANCE.md.

---

License

Apache License 2.0 — see LICENSE for details.