Airlock

A production-grade AI gateway built on Cloudflare Workers.

Airlock normalizes OpenAI, Anthropic, and Google Gemini APIs behind a unified request pipeline with first-class support for streaming, intelligent routing, key governance, and a built-in admin console.

中文文档

Features

Multi-protocol — Exposes OpenAI Chat Completions (/v1/chat/completions), OpenAI Responses (/v1/responses), and Anthropic Messages (/v1/messages) endpoints. Incoming requests are decoded into a canonical format, routed to the best provider, and re-encoded into the caller's protocol.
Multi-provider — Routes to OpenAI, Anthropic, and Google Gemini with provider-aware failover and cross-provider fallback.
Canonical data plane — All three protocols share a single normalized request/response pipeline. Cross-protocol translation (e.g. Anthropic client hitting an OpenAI model) works transparently.
Streaming — Full streaming support across all protocols and providers with stream reassembly, idle timeout detection, malformed SSE recovery, and protocol-correct mid-stream error events.
Intelligent routing — Five target selection strategies (weighted, lowest-cost, health-priority, priority, health-score) with per-route configuration. SLO-driven sliding-window circuit breakers protect against degraded providers.
Key governance — Static API keys and a dynamic key registry with full lifecycle management (create, rotate, archive, restore, revoke, staged rotation). Per-key request quotas, token quotas, and concurrency limits backed by Durable Objects.
Request shaping — Structured outbound request modification (headers, query params, JSON body injection) with HMAC-SHA256 request signing.
Observability — Structured request logging, telemetry event pipeline (Queue + Analytics Engine), per-route health metrics, and in-memory request statistics.
Admin API — Authenticated admin surface covering status, metrics, redacted config, raw config-store management, provider model discovery, routing health, and full key lifecycle management.
Console — React + HeroUI control plane UI with token login, metrics, key management, providers, routes, accounts, routing health, and an OpenAI Responses playground.
Production hardening — Request body size limits, Content-Type validation, CORS, timing-safe admin auth, per-IP rate limiting, client abort signal forwarding, empty stream detection, SSE buffer limits, and centralized error codes.
Cloudflare free-tier compatible — Runs on the Workers free plan by default. Scale mode unlocks higher throughput and full telemetry sampling.

Architecture

┌──────────────────────────────────────────────────────┐
│  Client (OpenAI / Anthropic / Gemini SDK)            │
└────────────────────┬─────────────────────────────────┘
                     │  /v1/chat/completions
                     │  /v1/responses
                     │  /v1/messages
┌────────────────────▼─────────────────────────────────┐
│  Gateway (Cloudflare Workers)                         │
│  ┌─────────┐  ┌───────────┐  ┌──────────┐           │
│  │Protocol │→│ Canonical  │→│ Routing  │            │
│  │Codec    │  │ Pipeline  │  │ Engine   │            │
│  └─────────┘  └───────────┘  └────┬─────┘           │
│                                    │                  │
│  ┌─────────────────────────────────▼──────────────┐  │
│  │  Provider Adapters (OpenAI / Anthropic / Gemini)│ │
│  └────────────────────────────────────────────────┘  │
│  ┌────────┐ ┌────────┐ ┌──────────┐ ┌────────────┐  │
│  │Govern. │ │Shaping │ │Telemetry │ │Admin API   │  │
│  └────────┘ └────────┘ └──────────┘ └────────────┘  │
└──────────────────────────────────────────────────────┘

Package Structure

Package	Purpose
`apps/gateway`	Cloudflare Worker entry point, HTTP routing, admin API
`apps/console`	React admin console (Cloudflare Pages)
`packages/protocols`	External protocol schemas and codecs (OpenAI Chat, OpenAI Responses, Anthropic Messages)
`packages/canonical`	Canonical request/response model, cross-protocol normalization, stream reassembly
`packages/providers`	Provider adapters (OpenAI, Anthropic, Gemini) with capability descriptors
`packages/routing`	Model routing, target selection, fallback, circuit breakers
`packages/governance`	Key authentication, quotas, revocation, dynamic registry, audit
`packages/request-shaping`	Structured outbound request modification and HMAC signing
`packages/telemetry`	Request event schema, queue consumer, Analytics Engine integration
`packages/shared`	GatewayError, error codes, environment utilities
`packages/testing`	Test factories and helpers

Quick Start

Prerequisites

Node.js >= 24
pnpm >= 11

Install

git clone <repo-url> airlock && cd airlock
pnpm install

Local Development

Copy the gateway environment file:

cp apps/gateway/.dev.vars.example apps/gateway/.dev.vars

Edit apps/gateway/.dev.vars. Minimal console-managed startup:

AIRLOCK_INTERNAL_ADMIN_TOKEN=dev-admin-token

With the Durable Object bindings from apps/gateway/wrangler.jsonc, this is enough to boot the gateway, connect from the console, and then configure providers, routes, caller keys, CORS, logging, limits, and policies in the UI.

Env-only fallback is still supported for deployments that intentionally avoid console-managed business config:

AIRLOCK_GATEWAY_API_KEYS=your-secret-key-here
AIRLOCK_PROVIDERS='[{"id":"openai-prod","type":"openai","apiKey":"sk-...","baseUrl":"https://api.openai.com/v1","defaultModel":"gpt-4.1-mini"}]'
AIRLOCK_MODEL_ALIASES='gpt-4.1-mini=openai-prod:gpt-4.1-mini'

Start the gateway dev server:

pnpm --filter @airlock/gateway dev

In a separate terminal, start the console:

pnpm --filter @airlock/console dev

Verify

# Run all checks
pnpm format && pnpm lint && pnpm typecheck && pnpm test && pnpm build && pnpm e2e && pnpm run audit

# Or individually
pnpm format        # Prettier check for tracked files
pnpm lint          # ESLint across all packages
pnpm typecheck     # TypeScript type checking
pnpm test          # Vitest across all packages
pnpm build         # Build all packages and apps
pnpm e2e           # Playwright console E2E tests
pnpm run audit     # Production dependency audit

Configuration

Airlock supports two configuration layers. The default production path is a minimal bootstrap plus console-managed business config:

Bootstrap env config for runtime bindings, admin auth, and safe fallbacks
Console overlay config from AIRLOCK_CONFIG_STORE for providers, routes, key policies, shaping, signing, model groups, limits, and feature flags
The dynamic key registry is enabled automatically when the AIRLOCK_GATEWAY_KEY_REGISTRY Durable Object binding exists

If AIRLOCK_CONFIG_STORE is absent, Airlock runs in env-only mode. See apps/gateway/.dev.vars.example for the full reference with descriptions and defaults.

Bootstrap Required

Variable	Description
`AIRLOCK_INTERNAL_ADMIN_TOKEN` or `AIRLOCK_INTERNAL_ADMIN_CREDENTIALS`	Admin API bootstrap auth

Business Config Required In Env-Only Mode

Variable	Description
`AIRLOCK_GATEWAY_API_KEYS`	API keys for caller auth unless registry-backed keys are used
`AIRLOCK_PROVIDERS`	JSON array of provider instances (`id`, `type`, `apiKey`, `baseUrl`)
`AIRLOCK_MODEL_ALIASES`	Env-side model routes targeting provider instance ids

Key Optional Variables

Variable	Description	Default
`AIRLOCK_MODE`	Operating mode: `free` or `scale`	`free`
`AIRLOCK_MODEL_ALIASES`	Env-side model routing (`external=provider:model`)	—
`AIRLOCK_MODEL_FALLBACKS`	Env-side fallback targets (JSON)	—
`AIRLOCK_MODEL_TARGET_SELECTION`	Env-side target selection strategy config (JSON)	—
`AIRLOCK_MODEL_KEY_POLICY`	Env-side route key access policy (JSON)	—
`AIRLOCK_MODEL_SHAPING`	Env-side outbound request shaping (JSON)	—
`AIRLOCK_MODEL_GROUPS`	Env-side model groups for key policies (JSON)	—
`AIRLOCK_PROVIDER_TIMEOUT_MS`	Upstream request timeout	`30000`
`AIRLOCK_PROVIDER_MAX_RETRIES`	Max cross-provider retries	`0`
`AIRLOCK_CORS_ORIGINS`	Env fallback CORS origins	—
`AIRLOCK_REQUEST_LOGGING`	Env fallback structured logging	`false`
`AIRLOCK_IP_RATE_LIMIT_POLICY`	Env fallback IP rate limit policy (JSON)	—
`AIRLOCK_CONFIG_STORE`	Console overlay config store binding	—
`AIRLOCK_GOOGLE_SUPER_ADMIN_EMAIL`	Map matching Google OAuth user to super admin	—

API Reference

Data Plane

Endpoint	Method	Description
`/v1/chat/completions`	POST	OpenAI Chat Completions protocol
`/v1/responses`	POST	OpenAI Responses protocol
`/v1/messages`	POST	Anthropic Messages protocol
`/v1/models`	GET	List available models
`/v1/models/:model`	GET	Get model details
`/v1/models/:provider/:model`	GET	Get provider-addressed model details

All /v1/* data plane endpoints require an API key via the Authorization: Bearer <key> header (or legacy x-api-key header).

Probes

Endpoint	Method	Description
`/healthz`	GET	Liveness probe
`/readyz`	GET	Readiness probe (provider verification)

Probe endpoints are unauthenticated so infrastructure health checks can call them directly.

Admin Plane

Admin endpoints require Authorization: Bearer <admin-token>. Legacy AIRLOCK_INTERNAL_ADMIN_TOKEN acts as a superuser token. Structured AIRLOCK_INTERNAL_ADMIN_CREDENTIALS enforce the scopes shown below.

Endpoint	Method	Scope	Description
`/_airlock/status`	GET	`status.read`	Gateway status and config fingerprint
`/_airlock/metrics`	GET	`metrics.read`	Request metrics (sliding window)
`/_airlock/config`	GET	`config.read`	Active configuration (secrets redacted)
`/_airlock/config/manage`	GET	`config.write`	Raw config-store snapshot
`/_airlock/config/manage/:section`	GET	`config.write`	Raw config-store section
`/_airlock/config/manage/:section`	PUT	`config.write`	Write config-store section
`/_airlock/config/manage/:section`	DELETE	`config.write`	Delete config-store section
`/_airlock/providers/fetch-models`	POST	`config.write`	Discover provider models with supplied credentials
`/_airlock/routing/health`	GET	`routing.read`	Per-route health and circuit breaker state
`/_airlock/keys`	GET	`keys.read`	List gateway keys
`/_airlock/keys`	POST	`keys.write`	Create a key
`/_airlock/keys`	PATCH	`keys.write`	Bulk update keys
`/_airlock/keys/bulk-create`	POST	`keys.write`	Bulk create keys
`/_airlock/keys/bulk-rotate`	POST	`keys.write`	Bulk rotate keys
`/_airlock/keys/bulk-delete`	POST	`keys.write`	Bulk delete keys
`/_airlock/keys/bulk-archive`	POST	`keys.write`	Bulk archive keys
`/_airlock/keys/bulk-restore`	POST	`keys.write`	Bulk restore keys
`/_airlock/keys/bulk-rotate/finalize`	POST	`keys.write`	Finalize bulk staged rotations
`/_airlock/keys/bulk-rotate/cancel`	POST	`keys.write`	Cancel bulk staged rotations
`/_airlock/keys/:id`	GET	`keys.read`	Get key details
`/_airlock/keys/:id`	PUT	`keys.write`	Update key metadata/policy
`/_airlock/keys/:id`	DELETE	`keys.write`	Delete a key
`/_airlock/keys/:id/rotate`	POST	`keys.write`	Rotate a key
`/_airlock/keys/:id/rotate/finalize`	POST	`keys.write`	Finalize staged key rotation
`/_airlock/keys/:id/rotate/cancel`	POST	`keys.write`	Cancel staged key rotation
`/_airlock/keys/:id/archive`	POST	`keys.write`	Archive a key
`/_airlock/keys/:id/restore`	POST	`keys.write`	Restore an archived key
`/_airlock/keys/:id/revocation`	GET	`keys.read`	Read key revocation status
`/_airlock/keys/:id/revocation`	POST	`keys.write`	Revoke a key
`/_airlock/keys/:id/revocation`	DELETE	`keys.write`	Clear key revocation
`/_airlock/keys/:id/status`	GET	`keys.read`	Key quota/status snapshot
`/_airlock/keys/:id/events`	GET	`keys.read`	Key audit events
`/_airlock/keys/operations/:operationId/events`	GET	`keys.read`	Operation-level key audit events
`/_airlock/keys/:id/registry`	GET	`keys.read`	Read registry override view
`/_airlock/keys/:id/registry`	PUT	`keys.write`	Set registry override
`/_airlock/keys/:id/registry`	DELETE	`keys.write`	Clear registry override

Deployment

Cloudflare Workers

The gateway deploys as a Cloudflare Worker. Required bindings:

Durable Objects — Key quota, token quota, concurrency, revocation, key registry, circuit breaker, IP rate limit
Queue — Telemetry event pipeline
Analytics Engine — Telemetry data storage

# Deploy to production
cd apps/gateway && pnpm wrangler deploy

# Or use the CI pipeline (pushes to main trigger automatic deployment)

Console (Cloudflare Pages)

cd apps/console && pnpm build
# Deploy the `dist/` directory to Cloudflare Pages

CI/CD

Two GitHub Actions workflows are included:

CI (ci.yml) — Runs on push and PR: lint, typecheck, test, build, format, production audit, and Playwright E2E
Deploy (deploy.yml) — Runs on push to main: verify + deploy to Cloudflare Workers (supports production and staging environments)

Model Routing Example

AIRLOCK_MODEL_ALIASES='gpt-4=openai-prod:gpt-4.1-mini,claude=anthropic-prod:claude-sonnet-4-5'
AIRLOCK_MODEL_FALLBACKS='{"gpt-4":["openai-compatible-a:gpt-4.1-mini","anthropic-prod:claude-sonnet-4-5"]}'
AIRLOCK_MODEL_TARGET_SELECTION='{"gpt-4":{"strategy":"weighted","weights":{"openai-prod:gpt-4.1-mini":10,"anthropic-prod:claude-sonnet-4-5":1}}}'
AIRLOCK_MODEL_KEY_POLICY='{"gpt-4":{"requiredKeyTier":"premium","requiredKeyTags":["chat"]}}'

Clients can then use gpt-4 or claude as model names — Airlock routes to the configured provider instance and falls back automatically on failures. Provider-addressed model IDs like openai-prod/gpt-4.1-mini are also accepted.

Testing

pnpm test           # Run all test suites
pnpm typecheck      # Type checking (tsgo)
pnpm build          # Build all packages
pnpm e2e            # Run console Playwright E2E tests
pnpm run audit      # Production dependency audit

The current verification suite covers 2000+ Vitest unit/integration tests plus Playwright E2E coverage for the console. Coverage spans protocols, canonical pipeline, providers, routing, governance, request shaping, telemetry, gateway integration, and console workflows.

Tech Stack

Runtime: Cloudflare Workers
HTTP: Hono
Validation: Zod
Language: TypeScript (strict, exactOptionalPropertyTypes, noUncheckedIndexedAccess)
Build: tsdown (libraries), wrangler (Worker), Vite (Console)
Frontend: React 19, TanStack Router, HeroUI, TailwindCSS 4
Testing: Vitest
Tooling: pnpm workspaces, ESLint, Prettier

License

MIT. See LICENSE.

Name		Name	Last commit message	Last commit date
Latest commit History 406 Commits
.github		.github
apps		apps
packages		packages
scripts		scripts
.gitignore		.gitignore
.prettierignore		.prettierignore
.prettierrc.json		.prettierrc.json
LICENSE		LICENSE
README.md		README.md
README.zh-CN.md		README.zh-CN.md
eslint.config.mjs		eslint.config.mjs
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
tsconfig.base.json		tsconfig.base.json
tsconfig.json		tsconfig.json

Navigation Menu

Uh oh!

Repository files navigation

Airlock

Features

Architecture

Package Structure

Quick Start

Prerequisites

Install

Local Development

Verify

Configuration

Bootstrap Required

Business Config Required In Env-Only Mode

Key Optional Variables

API Reference

Data Plane

Probes

Admin Plane

Deployment

Cloudflare Workers

Console (Cloudflare Pages)

CI/CD

Model Routing Example

Testing

Tech Stack

License

About

Releases

Packages

Used by

Contributors

Languages

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Airlock

Features

Architecture

Package Structure

Quick Start

Prerequisites

Install

Local Development

Verify

Configuration

Bootstrap Required

Business Config Required In Env-Only Mode

Key Optional Variables

API Reference

Data Plane

Probes

Admin Plane

Deployment

Cloudflare Workers

Console (Cloudflare Pages)

CI/CD

Model Routing Example

Testing

Tech Stack

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages