Airlock

Agentic Airlock

Identity & policy enforcement for AI agents — an open protocol that answers one question before any agent tool call runs: who is this agent, and is it authorized to do that? Cryptographic identity (Ed25519, W3C DID), deterministic policy decisions, and a signed, hash-chained receipt for every allow and every deny. Built on OAuth 2.1, with delegation chains and tamper-evident audit trails.

Works with Claude (Anthropic SDK integration included), OpenAI, LangChain, or your own agents — one identity & policy layer across all of them.

Registry: hosted registry in private beta — airlock.ing/registry

What's New in v1.0

OAuth 2.1 Authorization Server

• Full OAuth 2.1 server with private_key_jwt authentication (Ed25519)
• EdDSA-signed JWT access tokens with trust score claims
• RFC 8693 Token Exchange for delegation chains with scope narrowing
• Token introspection with live trust data, OIDC discovery, JWKS endpoints

Compliance Engine

• Agent inventory, risk classification (low/medium/high/critical), incident tracking
• Hash-chain integrity for tamper-evident compliance records
• Automated compliance report generation with regulatory framework mapping
• Bias detection for verification outcome patterns

Dual-Mode Identity Verification

• Orchestrator accepts both Ed25519 signatures and OAuth bearer tokens
• Backward-compatible — existing Ed25519 flows work unchanged

Semantic Challenge Deprecation

• LLM-based challenge disabled by default (now optional via pip install airlock-protocol[llm])
• Trust decisions based on cryptographic verification and behavioral scoring

See CHANGELOG.md for the full release history.

The Problem

AI agents now take real actions — they call tools, move data, run commands, execute transactions. A prompt injection or a misbehaving model turns a helpful agent into a confused deputy: a legitimate operation performed by the wrong agent, and no content guardrail will catch it. Guardrails filter content; they cannot distinguish an authorized tool call from an unauthorized one. There is no standard mechanism for verifying agent identity and enforcing authorization before the action runs.

The Solution

Airlock is an access-control layer that sits between the agent and its tools. Every tool call is checked — deterministically, whatever model is driving:

Agent ──tool call──> Airlock ──allowed──> Tool executes
                        │
                        ├── Identify   who is this agent? (Ed25519 / OAuth 2.1, verified not asserted)
                        ├── Decide     does policy allow it? (deterministic — same answer, any model)
                        ├── Enforce    allow or deny — before the tool ever runs
                        └── Receipt    signed, hash-chained record of every decision

The receipt chain is tamper-evident (Ed25519 signatures + SHA-256 hash chain) — it's the artifact you hand an auditor. OAuth 2.1 provides identity; Airlock adds trust scoring, delegation with scope narrowing (RFC 8693), and the enforcement + audit layer on top.

Architecture

                        +------------------------------------------+
                        |           Agentic Airlock                 |
                        |                                           |
  Agent A ---------->  |  [Gateway]  --->  EventBus                |
   (OAuth token or      |     |               |                     |
    Ed25519 handshake)  |     | ACK/NACK      v                    |
                        |     |         [Orchestrator]              |
                        |     |               |                     |
                        |     |         +-----+------+             |
                        |     |         v            v             |
                        |     |   ReputationStore  OAuth Server    |
                        |     |         |            |             |
                        |     |    trust score   token + claims    |
                        |     |         v                          |
                        |     |   TrustVerdict (VERIFIED /          |
                        |     |   REJECTED / DEFERRED)             |
                        |     |         |                          |
                        |     |         v                          |
                        |     |   Attestation + Compliance Audit   |
                        +-----+------------------------------------+

Quickstart

pip install airlock-protocol

# Verify an agent in 7 lines
python -c "
from airlock import AirlockClient
client = AirlockClient()  # defaults to api.airlock.ing
result = client.verify('did:key:z6MkhaXgBZDvotDkL5257faiztiGiC2QtKLGpbnnEGta2doK')
print(f'Verified: {result.verified}, Score: {result.trust_score}')
"

CLI

# Verify an agent from the command line
airlock verify did:key:z6Mk...

# Start a local gateway for development
airlock serve

# Scaffold a new Airlock-protected project
airlock init

Self-hosting

# Clone and run locally
git clone https://github.com/airlock-protocol/airlock.git
cd airlock
pip install -e ".[dev]"
python demo/run_demo.py       # 3-agent demo, no external services needed
python -m pytest tests/ -v    # 1,087 tests

Full Getting Started Guide

SDK Usage

from airlock import AirlockClient

# Default — routes through central Airlock registry (api.airlock.ing)
client = AirlockClient()
result = client.verify("did:key:z6Mk...")
if result.verified:
    print(f"Trusted: {result.agent_name}, Score: {result.trust_score}")

# Self-hosted — point to your own gateway
client = AirlockClient(gateway_url="http://localhost:8000")

# Async support
result = await client.averify("did:key:z6Mk...")

TypeScript client (airlock-client)

The npm workspace under sdks/typescript exposes the same REST operations via fetch (Node 18+). See sdks/typescript/README.md. Published PyPI name remains airlock-protocol (Python); the TS package is airlock-client on npm when released.

MCP adapter (airlock-mcp)

integrations/airlock-mcp is a stdio Model Context Protocol server that surfaces gateway tools (health, resolve, session, reputation, etc.) to MCP hosts. Build from repo root: npm install && npm run build:mcp.

When you publish: see RELEASING.md (PyPI OIDC, npm NPM_TOKEN, workflows).

Deploy (Docker)

• Docker Compose (gateway + Redis, persistent LanceDB volume): docs/deploy/docker.md
• Quick start: copy .env.example to .env, set AIRLOCK_GATEWAY_SEED_HEX, then docker compose up --build.

API Reference

Core Endpoints

OAuth 2.1 Endpoints

Compliance Endpoints

Operations Endpoints

Trust Scoring

Initial Score

New agents start at a neutral score of 0.50.

Routing Thresholds

Score Updates

Trust Tiers

Agents with 10+ interactions have a decay floor of 0.60 — established agents never drop back to fully unknown.

Half-Life Decay

Scores decay toward neutral (0.50) over time:

decayed = 0.5 + (score - 0.5) * 2^(-elapsed_days / half_life)

Decay half-life is tier-specific (see table above). An agent that stops interacting gradually becomes "unknown" rather than "suspect."

Project Structure

airlock-protocol/
├── airlock/
│   ├── config.py                  # Pydantic settings (env vars with AIRLOCK_ prefix)
│   ├── pow.py                     # Proof-of-Work (SHA-256 Hashcash / Argon2id)
│   ├── trust_jwt.py               # HS256 trust tokens for verified outcomes
│   ├── compliance/
│   │   ├── inventory.py           # Agent inventory registry
│   │   ├── risk_classifier.py     # Risk classification engine
│   │   ├── report_generator.py    # Compliance report generation
│   │   ├── incident.py            # Incident tracking with hash-chain integrity
│   │   ├── bias_detector.py       # Bias detection for verification patterns
│   │   ├── regulatory_mapper.py   # Regulatory framework principle mapping
│   │   └── schemas.py             # Compliance Pydantic models
│   ├── crypto/
│   │   ├── keys.py                # Ed25519 KeyPair + did:key encoding/decoding
│   │   ├── signing.py             # sign_model / verify_model + canonicalization
│   │   └── vc.py                  # W3C Verifiable Credential issue + validate
│   ├── engine/
│   │   ├── event_bus.py           # Typed async EventBus (asyncio.Queue backed)
│   │   ├── orchestrator.py        # LangGraph verification state machine
│   │   └── state.py               # SessionManager with TTL expiry
│   ├── gateway/
│   │   ├── app.py                 # FastAPI application factory + lifespan
│   │   ├── handlers.py            # Request handlers (dual-mode auth + event publish)
│   │   ├── routes.py              # Core protocol routes
│   │   ├── oauth_routes.py        # OAuth 2.1 endpoints
│   │   ├── compliance_routes.py   # Compliance endpoints
│   │   ├── revocation.py          # DID revocation store (sync + async + Redis)
│   │   └── rate_limit.py          # Per-IP + per-DID throttling
│   ├── oauth/
│   │   ├── server.py              # OAuth 2.1 authorization server
│   │   ├── grants/                # Client credentials + token exchange
│   │   ├── token_generator.py     # EdDSA JWT generation with trust claims
│   │   ├── token_validator.py     # JWT validation + delegation depth check
│   │   ├── introspection.py       # RFC 7662 token introspection
│   │   ├── discovery.py           # OIDC discovery + JWKS
│   │   ├── dependencies.py        # FastAPI Depends helpers
│   │   ├── registration.py        # RFC 7591 dynamic client registration
│   │   ├── models.py              # OAuth Pydantic models
│   │   ├── scopes.py              # Scope definitions + validation
│   │   └── store.py               # Client + token persistence
│   ├── reputation/
│   │   ├── scoring.py             # Tiered decay + verdict delta + floor protection
│   │   └── store.py               # LanceDB-backed TrustScore persistence
│   ├── rotation/                  # Key rotation with pre-rotation commitments
│   ├── schemas/                   # Pydantic models (identity, events, verdict, etc.)
│   ├── sdk/
│   │   ├── client.py              # AirlockClient (async httpx wrapper)
│   │   └── middleware.py          # AirlockMiddleware (protect decorator)
│   └── semantic/
│       ├── challenge.py           # LLM-backed challenge (optional, disabled by default)
│       └── fingerprint.py         # SimHash + SHA-256 bot detection
├── integrations/
│   └── airlock-mcp/               # MCP stdio server (gateway tools)
├── sdks/
│   └── typescript/                # npm package `airlock-client` (HTTP + types)
├── examples/                      # Agent scenarios + demos
└── tests/                         # 1,087 tests (unit, integration, property-based, security)

Design Principles

Environment Variables

All settings use the AIRLOCK_ prefix:

License

See LICENSE for details.

Author

Shivdeep Singh (@shivdeep1) — airlock.ing