Claude Code on Muapi

Environment variables

VariableValue
ANTHROPIC_BASE_URLhttps://api.muapi.ai/anthropic
ANTHROPIC_AUTH_TOKENYour Muapi API key (Claude Code sends it as a Bearer token)
ANTHROPIC_MODELOptional. An id from the models endpoint, e.g. claude-sonnet-5-5

Don't include /v1/messages in the base URL; Claude Code adds it.

Quick start (shell)

export ANTHROPIC_BASE_URL=https://api.muapi.ai/anthropic
export ANTHROPIC_AUTH_TOKEN="$MUAPI_API_KEY"
claude

Settings file

~/.claude/settings.json (or .claude/settings.local.json). Values here are literal, so prefer shell exports if you don't want the key in a file. Never commit credentials to a project-level .claude/settings.json.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.muapi.ai/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<your Muapi API key>"
  }
}

Models

curl -s -H "Authorization: Bearer $MUAPI_API_KEY" https://api.muapi.ai/anthropic/v1/models

Unrestricted (abliterated) models

The models endpoint also lists Muapi's tool-capable unrestricted models (for example qwen-3-8-27b-obliterated, mimo-v2-6-flash-abliterated, glm-5-3-abliterated, abliterated-model-large-v2). Claude Code can run on them: Muapi translates between Claude Code's protocol and the model's chat API for you.

export ANTHROPIC_BASE_URL=https://api.muapi.ai/anthropic
export ANTHROPIC_AUTH_TOKEN="$MUAPI_API_KEY"
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen-3-8-27b-obliterated
export ANTHROPIC_SMALL_FAST_MODEL=qwen-3-8-27b-obliterated
claude --model qwen-3-8-27b-obliterated

Set the two HAIKU/SMALL_FAST variables to the same model id. Claude Code sends background requests (titles, summaries) to a small "Haiku" model name, and without them those requests fail with a 404.

What to expect:

  • Quality varies by model. These are smaller open models; on multi-step coding tasks they can skip a step (for example forget to create one of several files). Check results, and try a larger model for harder work.
  • Cost grows with every turn. There is no prompt caching, so each turn is billed for the whole conversation so far. A typical turn costs roughly one cent to a few cents on the smaller models and more on the large ones; the Models endpoint and your dashboard show exact usage.
  • Output per response is capped at 16,384 tokens (some models lower), and the model's context window is smaller than Claude's, so very long sessions need /compact sooner.
  • Only models with tool calling are offered here; Claude Code cannot run without tools.
  • Image input works only on vision-capable models.

Verify

Run /status in Claude Code to confirm the base URL. Use claude --debug to see requests.

Notes

  • If Claude Code asks you to sign in to Anthropic, quit the terminal completely and reopen it from a shell where the variables are exported.
  • A 401 means the key or header is wrong; a 402 means your balance can't cover the request's worst case (the full prompt plus the maximum output is held first and the unused part is returned), so keep a buffer above the cost of one turn.