Skip to content

SETUP.md

Latest commit

 

History

History
25 lines (21 loc) · 2.46 KB

SETUP.md

File metadata and controls

25 lines (21 loc) · 2.46 KB
Raw

Gateway setup (checked against api.mor.org on 2026-10-27)

What the API offers

  • OpenAI-compatible base URL: https://api.mor.org/api/v1. GET /models lists 250+ models. They include Claude, GPT, Gemini, Grok, GLM, Qwen, Kimi and DeepSeek, :web variants with built-in web search, and embedding models.
  • Endpoints: /chat/completions, /embeddings and /audio/speech (text to speech). All need a key.
  • Not available: an Anthropic-style /messages endpoint, /responses, and image generation. Plan around these gaps.

Before you build

  1. One key per person from app.mor.org. Do not share keys. Keep them in a password manager and .env, never in git.
  2. Funding: buy credits at app.mor.org, or stake MOR to the API subnet (stakemor.com). Check your balance with the balance endpoint.
  3. Preflight: run ./morpheus/preflight.sh "<model A>" "<model B>". Then choose:
    • a demo model: fast, no long thinking, tool calls work. Thinking models can take more than a minute before the first token;
    • a fallback model, also on Morpheus;
    • an embedding model, if you need search or matching.
  4. Product code: one llm() function that uses the OpenAI SDK with baseURL = https://api.mor.org/api/v1. Put both model IDs in .env. Log the model and the latency of each call.
  5. Guard: run ./morpheus/no-other-vendors.sh as a pre-commit hook and in CI.
  6. Web research inside your product: use a :web model variant, so search also runs through Morpheus.

Coding agents (your build tools, not your product)

  • Agents that accept an OpenAI-compatible provider (for example OpenCode, Aider, Cursor) point straight at api.mor.org/api/v1. Refer to the integration guides.
  • Claude Code speaks /messages. It needs a local translator proxy, such as LiteLLM or claude-code-router, that turns /messages into /chat/completions, with ANTHROPIC_BASE_URL set to the proxy. Test tool calling and long sessions before you rely on it, because the translation can break tool use.
  • Decide early and say which one truthfully: all-Morpheus (every token, including the build agents) or product-on-Morpheus (the judged product runs 100% on Morpheus; the build tools do not).

On stage

Show the live call log: model ID, api.mor.org, latency. Keep a backup recording in case the venue Wi-Fi or a model fails.