Manifest is an LLM gateway that sits between your agents or apps and model providers. You connect API keys, consumer subscriptions, and local models, and every request goes through one OpenAI-compatible endpoint. Each query is routed to the model you configure, or you send "model": "auto" and the gateway selects one.
It connects to 300+ models through 32 built-in provider connections plus any custom OpenAI- or Anthropic-compatible endpoint. Provider catalogs are discovered when credentials are added. Alongside per-request API keys, 18 subscription flows reuse plans such as ChatGPT Plus/Pro/Team, Claude Max/Pro, Grok, Mistral Vibe, Kimi Coding Plan, MiniMax Coding Plan, GLM Coding Plan, GitHub Copilot, Kiro, ClinePass, OpenCode Go, and Ollama Cloud. Local models are served from Ollama, LM Studio, or llama.cpp.
The gateway records full request and response bodies for successes and errors, tracks spend per request with notifications and limits, and falls back to other models when a call fails, repairing malformed requests before the agent sees an error. It is available as a hosted cloud service or as a Docker image for self-hosting, with one-click templates for Railway, Render, DigitalOcean, AWS, GCP, Fly.io, Coolify, Easypanel, Heroku, Koyeb, and Apple Containers; each path provisions durable storage for request recordings.
Features
- Single endpoint: one OpenAI-compatible URL fronts every provider, with automatic model selection when the model is set to auto
- Provider breadth: OpenAI, Anthropic, Google, xAI, AWS Bedrock, Alibaba/Qwen, DeepSeek, Mistral, Moonshot, MiniMax, Xiaomi MiMo, Z.ai, OpenRouter, Cerebras, Fireworks, Groq, Hugging Face, NVIDIA NIM, and more
- Subscription routing: reuse existing consumer plans such as ChatGPT, Claude, Grok, Copilot, and Kiro instead of metered keys
- Local models: Ollama, LM Studio, and llama.cpp endpoints join the same routing table
- Custom providers: any /v1/chat/completions or /v1/messages endpoint can be registered
- Full body logs: request and response bodies are stored for both successful and failed calls
- Cost tracking: per-request spend accounting with notifications and limits
- Fallback and repair: failed queries retry on alternative models and broken requests are rewritten before the caller sees them
- Deployment: Docker image with templates for Railway, Render, AWS, GCP, DigitalOcean, Fly.io, and others, or the hosted cloud
