Manifest LLM Gateway is an open-source LLM gateway for AI agents and apps. It puts API keys, subscription plans and local models behind one OpenAI-compatible endpoint and routes each query to a model, so an agent is not tied to a single provider. Sending the model name auto lets Manifest pick the model, and routes can carry their own model and parameter settings without code changes.
It connects to more than 300 models through 35 built-in provider connections, plus any custom OpenAI- or Anthropic-compatible endpoint. Credentials can be a provider API key, one of 18 subscription sign-in flows (such as ChatGPT, Claude, Gemini, Grok, Kimi, GLM and GitHub Copilot plans), or a local server such as Ollama, LM Studio or llama.cpp. Provider model catalogs are discovered when credentials are connected.
When a query fails, Manifest falls back to another model or provider, and an autofix step repairs malformed requests before the error reaches the agent. Every call is logged with its full request and response body, including errors, and spend is tracked per agent, key and provider with notifications and limits. Manifest runs as a hosted cloud version or self-hosted from a Docker image, with deployment guides for Railway, Render, DigitalOcean, AWS, GCP, Fly.io, Coolify, Easypanel, Heroku, Koyeb and Apple Containers.
Features
- Single endpoint: one OpenAI-compatible endpoint for every connected provider, with automatic model selection on request
- Custom routing: route by API keys, subscriptions, local models or custom providers, with per-route model and parameter settings
- Provider coverage: 35 built-in connections including OpenAI, Anthropic, Google, Vertex AI, xAI, AWS Bedrock, Qwen, DeepSeek, Mistral, Moonshot, MiniMax, Xiaomi MiMo, Z.ai, OpenRouter, Groq and NVIDIA NIM
- Subscription flows: 18 sign-in flows that reuse plans such as ChatGPT, Claude, Gemini, Grok, Kimi, GLM, Mistral Vibe, GitHub Copilot and Kiro
- Local models: routes to Ollama, LM Studio, llama.cpp or any local server, deciding what goes to the cloud and what stays on the machine
- Fallback and autofix: retries on another model or provider when a query fails, and repairs broken requests before the agent sees the error
- Full body logs: complete request and response bodies for successful and failed calls, kept in durable recording storage
- Cost tracking: spend broken down by agent, key and provider, with notifications and limits
- Deployment options: a hosted cloud version, a self-hosted Docker image, and one-click templates for Railway, Render, DigitalOcean, AWS and GCP
- n8n node: a community node that connects n8n workflows to the gateway
