Cody Gateway

An AI API gateway with a web console, for Codex, Claude Code and other OpenAI- or Anthropic-compatible clients. It runs on Cloudflare Workers, as a Node.js server, or on Vercel.

• Manage AI Gateway, Antigravity, Codex and Claude providers, credentials, client keys and model aliases in the console.
• Route traffic through SOCKS5 proxy groups.
• Track usage, costs and request details, with custom pricing by provider and model.

Keep your deployment’s encryption key safe and unchanged. Never commit it or files containing secrets, such as .env and .dev.vars.

All platforms run the same gateway and console. Every build requires Node.js 24 or newer.

libSQL works with Turso or a self-hosted sqld.

Inference upload limits apply separately before and after decompression. The effective upload limit is the smaller of the operation budget and the deployment ceiling; Workers use a 16 MiB buffering budget to leave memory for decoded text, parsed JSON and translated requests within their 128 MB isolate limit. Vercel also narrows management requests to 4.5 MB. Reverse proxies and upstream APIs may impose smaller limits. See the Vercel Function limits and Cloudflare WebSocket limits. Response inspection and usage observation have independent 16 MiB budgets; reaching an observation limit does not truncate a forwarded response.

HTTP inference and context-management requests run in a dedicated HttpExecution Durable Object per request. The entry Worker forwards request and response streams without parsing them; authentication, routing, provider adapters and usage observation run inside the object. Existing account, health and session objects still coordinate requests. WebSocket upgrades keep their existing path. Each executor accepts one request. Failed dispatches cancel that executor without resending inference. Client disconnects cancel the executor through a separate RPC, including before upstream headers arrive or while response writes are blocked. The executor waits for pending usage journal writes before completing cancellation. This moves inference CPU work out of the entry Worker's budget, but active HTTP streams consume Durable Object duration quota. Deployment adds DO migration v10; the executor does not write request bodies or tokens to its own storage.

npm install
npm run dev:cloudflare

Open http://127.0.0.1:5173/console/ for the console with hot reload. The gateway endpoint is http://127.0.0.1:8787/v1.

The first run creates a local database and .dev.vars with an encryption key. Keep this key: existing data depends on it.

1. Create the Cloudflare resources:

   npx wrangler login
   npx wrangler kv namespace create CODY_CONFIG_KV
   npx wrangler d1 create cody
   npx wrangler queues create cody-usage
   npx wrangler queues create cody-usage-dlq

2. Put the returned KV namespace and D1 database IDs in wrangler.jsonc.

3. Bind your domain to the Worker and protect the console with Cloudflare Access: create a self-hosted application for the path console/* and allow your administrator emails. In wrangler.jsonc, set ACCESS_TEAM_DOMAIN to https://your-team.cloudflareaccess.com, ACCESS_AUD to the application's AUD tag, and keep ADMIN_LOCAL_DEV=false.

4. Store an encryption key and deploy:

   openssl rand -base64 32
   npx wrangler secret put CONFIG_ENCRYPTION_KEY
   npm run deploy:cloudflare

npm run deploy:cloudflare builds the console, applies database migrations and deploys the Worker.

For automatic deployments, connect the repository to Cloudflare Workers Builds and use npm run deploy:cloudflare as the deploy command. The API token needs D1 Edit permission in addition to the Worker permissions.

You need a Redis server. SQLite is the default database; use PostgreSQL or libSQL when running more than one instance.

npm ci
cp .env.example .env
openssl rand -base64 32

Edit .env:

• CONFIG_ENCRYPTION_KEY: the generated key.
• REDIS_URL: your Redis server.
• DATABASE_URL: keep the default sqlite:./data/cody.sqlite, or use postgres://user:password@host:5432/cody or libsql://your-database.turso.io?authToken=your-token.

npm run build:node
npm run start:node

Open http://127.0.0.1:8787/console/. The gateway endpoint is http://127.0.0.1:8787/v1. The default ADMIN_AUTH_MODE=local signs you in automatically and only works while the server listens on 127.0.0.1.

For development with hot reload, use npm run dev:node and open http://127.0.0.1:5173/console/. Keep HOST=127.0.0.1 and PORT=8787 in .env.

Before exposing the server, switch console sign-in to OIDC (see Console sign-in) and set HOST=0.0.0.0.

With Docker:

docker build -t cody .
docker run -d -p 8787:8787 -v cody-data:/data --env-file .env -e DATABASE_URL=sqlite:/data/cody.sqlite cody

Inside the container, 127.0.0.1 refers to the container itself, so point REDIS_URL at a Redis address the container can reach.

Without Docker, use npm run build:node followed by npm run start:node. To update, run npm run deploy:node, then restart the server.

Behind a reverse proxy that terminates TLS, preserve the public Host header and set X-Forwarded-Proto: https.

You need a PostgreSQL or libSQL database and a Redis server that Vercel can reach. Place them near your function's region.

1. Import the repository into Vercel. Select Node.js 24 and leave Root Directory empty.
2. Add the environment variables:
   • CONFIG_ENCRYPTION_KEY: generate with openssl rand -base64 32.
   • DATABASE_URL: your PostgreSQL or libSQL connection string.
   • REDIS_URL: your Redis server.
   • CRON_SECRET: a random string of at least 16 characters.
   • Console sign-in variables (see Console sign-in).
3. Deploy. Vercel runs npm run deploy:vercel (set in vercel.json), which builds and applies database migrations before release; a failed migration stops the deployment.

Open /console/ on your deployment URL. Keep each preview environment on its own database and REDIS_PREFIX, separate from production.

Vercel supports HTTP only, with a default request timeout of 300 seconds.

For browser sign-in, use any OIDC provider:

ADMIN_AUTH_MODE=oidc
ADMIN_OIDC_ISSUER=https://identity.example.com
ADMIN_OIDC_CLIENT_ID=cody
ADMIN_OIDC_CLIENT_SECRET=your-client-secret
ADMIN_OIDC_ALLOWED_EMAILS=admin@example.com,another-admin@example.com

Register https://your-gateway/console/auth/callback as the redirect URL. The provider must return verified email addresses.

Other modes:

• ADMIN_AUTH_MODE=access: behind Cloudflare Access, for example through a Cloudflare Tunnel. Set ACCESS_TEAM_DOMAIN and ACCESS_AUD.
• ADMIN_AUTH_MODE=token: API access only, with ADMIN_TOKEN of at least 16 characters sent as Authorization: Bearer ….

Use a client key created in the console and a model enabled in your published configuration.

Other clients can use the gateway's OpenAI or Anthropic endpoints with a client key, sent as Authorization: Bearer or x-api-key.

Add a provider to ~/.codex/config.toml, replacing base_url with your gateway URL and choosing a configured model:

model = "gpt-5.6-sol"
model_provider = "gateway"

[model_providers.gateway]
name = "Gateway"
base_url = "https://cody.example.com/v1"
wire_api = "responses"
http_headers = { "x-openai-actor-authorization" = "cody" }

[model_providers.gateway.auth]
command = "printenv"
args = ["OPENAI_API_KEY"]
timeout_ms = 5000
refresh_interval_ms = 300000

export OPENAI_API_KEY="your-gateway-client-key"
codex

export ANTHROPIC_BASE_URL="https://cody.example.com"
export ANTHROPIC_API_KEY="your-gateway-client-key"
claude --model "your-configured-model"

Use npm run data:transfer -- --help to copy business database data and encrypted OAuth accounts between Cloudflare D1, SQLite, PostgreSQL and libSQL. D1 uses Wrangler SQL-file export/import; standard databases use bulk reads and inserts. Runtime caches, health/cooldown state and unrecorded queue/outbox usage are excluded. Both deployments must use the same code, applied schema and CONFIG_ENCRYPTION_KEY; the target must be fresh.

Stop requests and background jobs on both sides before confirming the CLI prompt. Progress is kept locally for --resume or --abort; no database migration or runtime maintenance lock is added. D1 transfers also require the sqlite3 CLI. Start a new task for this workflow; older task files are not supported. See the transfer guide for setup and cutover.

npm test
npm run typecheck
npm run lint
npm run format:check