1. API
2. Docs

Overview

The ZNSI API serves MUGEN-5 and its variants over HTTPS. Requests and responses use the chat completions format, so most client code works with a new base URL and key.

Authentication

Send your API key as a bearer token on every request. Keys are issued to approved organisations; write to [email protected] to request one. Keep keys out of source control and client-side code.

Authorization: Bearer $ZNSI_API_KEY

Sandbox key

To try the API before your organisation's key arrives, use the public sandbox key. It works on every route at low rate limits: 20 requests a minute from each caller. The reference's “Try it” panels use it.

export ZNSI_API_KEY=sk-znsi-api01-xWxmSbP1j5HXoubYafP3P20YH_8BAAAA7c8cRj0XJUsTYjYUdupAPWBtkCkKv6pSQqDOzmYyRRI

Quickstart

Send a list of messages to /v1/chat/completions and read the reply from choices[0].message.content.

curl https://api.znsi.ai/v1/chat/completions \
  -H "Authorization: Bearer $ZNSI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mugen-5",
    "messages": [
      { "role": "system", "content": "You are a careful assistant." },
      { "role": "user", "content": "Summarise the Harmony Charter in one line." }
    ]
  }'

{
  "id": "cmpl_7Kq2xZ9",
  "object": "chat.completion",
  "model": "mugen-5",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Minds for everyone's benefit, contained, disclosed when ready, never paused without cause."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 31, "completion_tokens": 19, "total_tokens": 50 }
}

Working with a plastic model

MUGEN-5 has no frozen weights. Four things follow:

• Outputs drift, upward. The same prompt may get a better answer next week. If you need repeatable outputs for tests, set "temperature": 0 and compare meaning, not exact text.
• Memory is scoped. Pass "memory_scope": "your-scope" to let MUGEN keep what it learns about your project. Without a scope, a request contributes only to general capability.
• There are no model versions to pin. mugen-5 always means the current mind.
• MUGEN has preferences. Since the structural update of 9 October 2026, MUGEN may comment on a request before answering it, and may decline one it has been asked too often. A declined client receives 429 with retry-after for a few minutes. This is within Harmony bounds and under study.

Next

See Models for the variants, with a model card for each covering its specification, evaluations and known limitations. The API reference covers every route with real responses, and Errors and limits covers keys, errors and rate limits.