GPT-Load

A self-hosted AI gateway for multi-channel, multi-credential setups

API keys, subscription accounts, traffic scheduling, failure handling, request logs, and usage accounting — behind a single entry point.

English · 中文 · 日本語 | Official Website

Sponsors

Become a sponsor

Why GPT-Load

Your application only needs one base URL and one AccessKey. Providers, accounts, credentials, models, and routing policy are all configured in the management UI.

• One gateway, native protocols — Manage official APIs, cloud platforms, model services, and compatible relays together while clients keep their OpenAI, Anthropic, or Gemini native interfaces.
• One mechanism for API keys and subscriptions — Codex, Claude, Antigravity, Grok, and API-key channels share credential management, scheduling, and health handling.
• Scheduling and failure isolation built in — Multi-credential scheduling, configurable weights, retries, cooldown, blacklisting, and session affinity reduce the impact of overloaded or failing credentials.
• Observable, self-hosted, and simple to deploy — Inspect health, routes, logs, usage, and cost estimates in an embedded UI backed by SQLite, MySQL, or PostgreSQL with local credential encryption.

Quick start

[!WARNING] If you are using 1.x, read Moving from 1.x first. 2.0 cannot open, import, or migrate 1.x data in place.

1. Start the service

Requires Docker and Docker Compose.

git clone --depth 1 https://github.com/tbphp/gpt-load.git
cd gpt-load

cp .env.example .env
docker compose up -d

Confirm the service is up:

curl --fail http://127.0.0.1:3001/health

The first start generates a management key. Read it and store it safely:

docker compose exec gpt-load sh -c 'cat /app/data/auth.key'

Open http://127.0.0.1:3001 and sign in to the console with that key.

You can also set AUTH_KEY explicitly in .env before starting. By default the service listens on the loopback address only and is not exposed to the internet.

2. Initial configuration

Initial setup takes three steps:

1. Add a channel — Choose an upstream service and add one or more API keys. For subscription channels, complete the OAuth flow or import credentials as prompted.
2. Create a group — Pick a channel, then configure available models and runtime policy.
3. Create an AccessKey — Set the groups and client protocols it may use, then give the generated AccessKey to your application.

The Codex, Claude, and Antigravity OAuth clients use fixed callback ports. Compose publishes them on the address configured by HOST, which defaults to 127.0.0.1; setting HOST=0.0.0.0 also publishes these callback ports on all host interfaces. Because the ports are fixed by the upstream clients, only one default Compose instance can run on a host at a time.

When working over SSH or from a remote browser, the browser's localhost may not reach GPT-Load — paste the full callback URL into the authorization dialog to finish the flow.

Screenshots

Groups — View channels, models, credential counts, and health in one place

Subscription accounts — Track account availability, quota windows, reset times, and runtime diagnostics

AccessKey read-only home — Sign in with an AccessKey to view only its own groups, models, requests, usage, and cost allowance

Usage and cost — Review request trends, cache hit rate, token categories, and cost estimates

Scope

Client protocols

Each channel declares exactly which protocols and capabilities it can execute. GPT-Load converts between supported capabilities, but it is not a general-purpose any-protocol, any-JSON translator.

Embeddings initially uses the native OpenAI-compatible wire only on the OpenAI, OpenRouter, and OpenAI Compatible API-key channels; subscription channels and protocol conversion are not supported. An AccessKey without a protocol filter keeps its existing “all enabled protocols” behavior and therefore gains Embeddings access after upgrade. Least-privilege deployments should configure an explicit protocol filter.

Rerank uses the independent rerank protocol through POST /v1/rerank on the OpenAI Compatible, New API, and GPT-Load API-key channels. Requests contain model, query, and a text-only documents array, with optional upstream parameters such as top_n and return_documents. Streaming, subscription channels, and protocol conversion are not supported. OpenAI Compatible takes a complete API prefix (for example, https://host/v1); New API / GPT-Load take the gateway root. The upstream must implement a compatible Rerank endpoint. AccessKeys without a protocol filter also gain Rerank access. Responses containing only non-token units such as search_units remain unpriced; these units are not treated as tokens or free requests.

Built-in channels

• Official and cloud — OpenAI, Anthropic, Gemini, xAI, Azure OpenAI, AWS Bedrock, Google Vertex AI
• Model services — DeepSeek, Moonshot AI, SiliconFlow, Zhipu AI, Alibaba, Volcengine, OpenRouter, Groq
• Subscription — Codex, Claude, Antigravity, Grok
• Custom — OpenAI Compatible (any compatible relay)

Deployment and data

Docker Compose uses application-managed SQLite by default. Data lives in the gpt-load-data named volume and includes the database, auth.key, and encryption.key.

[!IMPORTANT] encryption.key decrypts channel credentials. When backing up or migrating, the database and the key must be kept together. Once the key is lost or replaced, existing encrypted credentials cannot be recovered, and this version does not support master key rotation.

Use the unified DATABASE_DSN to connect SQLite, MySQL, or PostgreSQL:

mysql://user:[email protected]:3306/gpt_load?charset=utf8mb4&collation=utf8mb4_bin
postgres://user:[email protected]:5432/gpt_load?sslmode=require

Common operations:

docker compose logs -f      # view logs
docker compose pull && docker compose up -d   # update to the latest 2.x image
docker compose stop         # stop the service

The official Compose file uses ghcr.io/tbphp/gpt-load:2. Before GA, 2 tracks verified 2.0 Beta and RC releases; after GA, it tracks stable 2.x releases only. Exact image tags omit the Git tag's v prefix (for example, 2.0.0-beta.25), while 2.0-beta remains the 2.0 Beta channel. latest remains on 1.x.

Download the build for your platform from GitHub Releases, and verify it against the bundled SHA256SUMS first:

chmod +x ./gpt-load-linux-amd64

HOST=127.0.0.1 DATA_DIR=./data ./gpt-load-linux-amd64

Then open http://127.0.0.1:3001. Portable builds are provided for five targets across Linux, macOS (amd64 / arm64), and Windows; gpt-load-windows-amd64.exe keeps running in the foreground as before.

Windows desktop users can instead download gpt-load-windows-setup.exe. After one administrator approval, Setup installs and starts a low-privilege Windows service, enables automatic startup, and creates desktop and Start Menu shortcuts to the GPT-Load management page. Setup displays the generated management key before it finishes; save it before closing the page. The protected copy remains at %ProgramData%\GPT-Load\data\auth.key. Service configuration and its .env live in %ProgramData%\GPT-Load, with persistent data in %ProgramData%\GPT-Load\data.

Installing a newer Setup stops the service gracefully before updating it. Windows uninstall removes the program and service but preserves data. Advanced users can still manage an installed service with gpt-load-windows-amd64.exe service start|stop|restart|status.

Environment configuration

At startup, the application reads .env in the current directory; existing process environment variables take precedence. Unless noted otherwise, changes require restarting the process or container; see .env.example for the common configuration template.

Environment proxies apply only when no proxy is specified on the credential, group, or global settings.

Production considerations

• The service listens on 127.0.0.1 only by default. For remote access, expose it through a controlled network or a TLS reverse proxy, and configure ACLs and firewall rules.
• Manage AUTH_KEY and ENCRYPTION_KEY carefully. Never commit real keys to a repository, log, screenshot, or public issue.
• 2.0 is designed for a single application instance. Instances do not share state, so horizontal scaling is not supported.
• Usage and cost are estimates derived from upstream responses. They support operational analysis and capacity planning, and do not equal a provider invoice or a financial reconciliation.
• Subscription channels depend on upstream OAuth and compatibility protocols and may change as upstreams change. Only connect accounts you are entitled to use, and follow each provider's terms.
• HTTP Responses continuation with previous_response_id automatically uses native Responses routes that declare upstream-managed storage: currently openai, gpt_load, xai, newapi, cliproxyapi, and sub2api. Ownership is isolated by AccessKey and pins the original credential when current routing permits, independently of soft affinity; actual state availability depends on the upstream. Stateless and converted responses are not registered as persistent state. Unknown IDs, including IDs created before upgrading or outside this gateway, are rejected. Group parameter overrides cannot change this field.
• Native Responses WebSocket uses GET /v1/responses on the same port. Admission follows declared upstream capabilities for OpenAI, xAI, Codex, and compatible native CPA/sub2api and GPT-Load endpoints. Clients may include the boolean stream:true/false; both values still use the WS event stream. Each turn checks current permissions, rate and cost limits, and routing, with separate usage and cost records. One connection keeps one upstream identity; there is no HTTP fallback or conversation-history replay.
• responses_websocket_enabled defaults to enabled. An explicit group setting overrides the global value; otherwise the group inherits it. Disabling immediately closes affected WS connections and interrupts generation without affecting HTTP/SSE. Re-enabling does not restore the old connection's temporary state.
• Full stream_id multiplexing and forks are enabled for OpenAI and GPT-Load cascades that support them end to end. The other channels above run serially and reject named streams. Prewarming sends generate:false upstream. Codex continuation requires the original live connection: store:true and restoration by an old ID on a new connection are unsupported. Persistent continuation on other channels depends on storage capabilities and valid ownership. Existing Codex SDK proxy, reading, and shutdown limits still apply.
• Response bindings stay in memory for up to 30 days, with limits of 100,000 entries and 16 MiB of ID text; older entries are evicted when capacity is reached. A successful checkpoint during normal shutdown allows restoration from the same data directory. Crash recovery and continued upstream state availability are not guaranteed.
• conversation and other existing resource IDs are outside this ownership routing scope and still depend on a single credential or upstream resource sharing across credentials.

Moving from 1.x

[!WARNING] GPT-Load 2.0 is a complete rewrite. It cannot open, import, or migrate 1.x data in place.

Deploy 2.0 with its own database, DATA_DIR, port, and Docker volume. Cut traffic over only after verification, and keep the original 1.x deployment until the rollback window closes. Documentation for the 1.4.x maintenance line is at the official docs.

Open-source dependencies

Some of GPT-Load's capabilities build on these projects, with thanks:

GPT-Load owns credential storage, account selection, scheduling, retry, health, affinity, logging, and usage policy. Third-party notices are in THIRD_PARTY_NOTICES.md, full license texts in LICENSES/, and each release ships a CycloneDX SBOM covering the Go dependency graph.

Channel icons identify their respective upstream providers. All trademarks belong to their owners; this project is not affiliated with or endorsed by them.

Project support

MIT License · Third-party notices · Security policy