LLM Gateway Lite

Give Cursor BYOK every model LiteLLM supports.

Cursor's Bring Your Own Key path speaks OpenAI. LiteLLM already speaks 100+ providers. This repo is the missing piece: a personal Docker gateway that sits between Cursor and those providers, so Ask, Plan, and Agent can use OpenAI, Anthropic, Gemini, xAI/Grok, Azure, Bedrock, Ollama, OpenRouter, and any other model LiteLLM can route.

English | 简体中文 | 日本語

Why this exists

Cursor BYOK is one OpenAI-compatible Base URL plus your key. LiteLLM is the widest OpenAI-compatible translator in the ecosystem. Point Cursor at this gateway and every LiteLLM-supported model becomes a custom Cursor model:

1. Add the upstream in the LiteLLM UI with your own provider key.
2. Issue a Virtual Key for Cursor.
3. Set Cursor's Override OpenAI Base URL to https://your-domain/cursor.
4. Chat with that public model name in Ask, Plan, and Agent.

You keep the keys. You keep the spend. Cursor keeps the IDE. LiteLLM does the protocol work.

How it works

flowchart LR
  Cursor["Cursor IDE<br/>BYOK"] -->|"HTTPS /cursor"| GW["LLM Gateway Lite<br/>LiteLLM Proxy"]
  GW --> OpenAI
  GW --> Anthropic
  GW --> Gemini
  GW --> Grok["xAI / Grok"]
  GW --> Cloud["Azure / Bedrock / Vertex"]
  GW --> Local["Ollama / vLLM"]
  GW --> More["100+ LiteLLM providers"]

Loading

The local stack is three containers:

Service	Role
litellm	Official LiteLLM UI, OpenAI-compatible /v1, Cursor /cursor
db	PostgreSQL for models, credentials, Virtual Keys, budgets, usage
redis	Routing coordination and fallbacks

The image pins LiteLLM v1.100.0 (Agent mode needs v1.97.0+). Models live in the database (store_model_in_db: true). The custom hook only enforces the upstream call-id length compatibility rule.

Cursor coverage follows what Cursor enables for custom API keys. Use a public HTTPS URL Cursor's servers can reach; localhost is for running and testing the gateway itself.

Quick start

Copy .env.example to .env and fill in random secrets. scripts/init.* generate the Master Key, Salt Key, Postgres password, and local URL. Local Compose also needs UI_PASSWORD and REDIS_PASSWORD.

Windows PowerShell:

Copy-Item .env.example .env
notepad .env
docker compose up -d --build

Linux or macOS:

cp .env.example .env
# edit .env with random secrets
docker compose up -d --build

Use	URL
Admin UI	http://localhost:3029/ui/
Readiness	http://localhost:3029/health/readiness
OpenAI-compatible API	http://localhost:3029/v1/
Cursor Base URL	http://localhost:3029/cursor

Sign in with UI_USERNAME (default admin) and UI_PASSWORD. LITELLM_MASTER_KEY is the proxy admin API key. Cursor uses a Virtual Key.

Point Cursor at it

In Cursor Settings → Models:

1. Enable OpenAI API Key and paste a LiteLLM Virtual Key.
2. Enable Override OpenAI Base URL.
3. Set Base URL to https://your-domain/cursor (local smoke test: http://localhost:3029/cursor).
4. Add the LiteLLM Public Model Name.

Base URL:  https://your-domain/cursor
API Key:   LiteLLM Virtual Key
Model:     your-public-model-name

That /cursor path is the official LiteLLM Cursor integration. If Cursor already ships a model under the same name, register a distinct public alias in LiteLLM and use that alias in Cursor.

Add any LiteLLM model

Open Models + Endpoints in the UI:

Field	Meaning
Public Model Name	Name Cursor and other clients call
LiteLLM Model Name	Provider, protocol, and upstream model, e.g. anthropic/claude-sonnet-4-6 or openai/responses/grok-4.6
API Base	Upstream root, e.g. https://api.example.com/v1
API Key	Your provider key
RPM / TPM	Optional per-deployment limits

The same public name can attach multiple deployments. LiteLLM routes and can fail over between them.

Then create a Virtual Key, grant it those public names (or *), and paste that key into Cursor. Keep Master Key, Salt Key, and upstream keys on the gateway only.

Deploy on Rainyun

Production in this project targets Rainyun RCA (Rain Cloud Apps): import rainyun-compose.yml, put HTTPS in front, and point Cursor at https://your-domain/cursor.

New to Rainyun? Open https://www.rainyun.com/Nzc5MDEw_, create a project with at least 2 GB RAM, then follow Configuration & troubleshooting for Compose import, secrets, website proxy, backup, and upgrades. Also: 简体中文 · 日本語.

The Rainyun template is LiteLLM + PostgreSQL. Local Compose also runs Redis. Add Redis in the cloud when you scale replicas or need shared rate-limit and routing state.

What you get

• Cursor Ask, Plan, and Agent through LiteLLM's /cursor entry
• Every provider LiteLLM already knows, behind one OpenAI-compatible URL
• Official LiteLLM admin UI, Virtual Keys, budgets, and usage
• Upstream-compatible tool-call ID length normalization
• Local Docker Compose and a Rainyun RCA template

Docs

Rainyun import, resource sizing, backup, upgrades, security checklist, and longer troubleshooting:

Configuration & troubleshooting · 简体中文 · 日本語

node tests/check-static.mjs
docker compose config --quiet
docker compose -f rainyun-compose.yml config --no-interpolate --quiet
docker build -t llm-gateway-lite ./litellm

License

MIT