Use with Claude Code

Claude Code speaks the Anthropic Messages API and lets you redirect it at any compatible endpoint with a couple of environment variables. Otari exposes that surface (POST /v1/messages and POST /v1/messages/count_tokens) in both standalone and hybrid modes, so you can route Claude Code through Otari to get budgets, usage tracking, and traces without changing how you use the CLI.

Quick start

export ANTHROPIC_BASE_URL="https://api.otari.ai"   # your Otari base URL, no /v1
export ANTHROPIC_AUTH_TOKEN="tk_your_otari_token"  # sent as Authorization: Bearer
export ANTHROPIC_MODEL="anthropic:claude-sonnet-4-6"
claude

Claude Code appends /v1/messages and /v1/messages/count_tokens to ANTHROPIC_BASE_URL itself, so the base URL must be the Otari root (for local development, http://localhost:8000).

Standalone mode: allow Claude Code's user_id

Claude Code attaches its own metadata.user_id to every request. In standalone mode Otari binds spend to the API key's own user and, by default, rejects a request that names a different user (403 permission_error). Set reject_user_mismatch: false in your config so Claude Code's user_id is ignored and spend is bound to the key. (Hybrid mode authenticates via the user token and does not compare metadata.user_id, so this setting does not apply there.)

Use ANTHROPIC_AUTH_TOKEN (not ANTHROPIC_API_KEY): it is sent as Authorization: Bearer <token>, which is the scheme Otari accepts for both standalone API keys and connected user tokens. ANTHROPIC_API_KEY sends an x-api-key header instead, which Otari does not read.

settings.json

The same configuration works in ~/.claude/settings.json (or a project-level .claude/settings.json):

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.otari.ai",
    "ANTHROPIC_AUTH_TOKEN": "tk_your_otari_token",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "anthropic:claude-opus-4-8",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "anthropic:claude-sonnet-4-6",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "anthropic:claude-haiku-4-5"
  }
}

You are not limited to Claude models

Claude Code speaks the Anthropic wire format, but the model field is just a string Otari forwards to any-llm. any-llm translates the Messages format to and from each provider's native API, so any model in the catalog works — OpenAI, Mistral, Mozilla.ai inference models, Moonshot, and so on — not just Anthropic's Opus/Sonnet/Haiku.

Set the model strings to whatever your deployment expects:

• Connected to otari.ai: use mzai:<catalog-id> for a managed open-weight model (e.g. mzai:moonshotai/Kimi-K2.6), or provider/model for one of your own provider keys (e.g. openai/gpt-4o, anthropic/claude-sonnet-4-6). An mzai: prefix selects the managed catalog, so adding it to a proprietary model routes it there instead of your key and it will not resolve.
• Standalone: use provider:model, e.g. openai:gpt-4o, mistral:mistral-large-latest, or anthropic:claude-sonnet-4-6.

Claude Code uses two tiers you can point independently: a primary model for the agent loop (ANTHROPIC_MODEL, or the ANTHROPIC_DEFAULT_*_MODEL tier the opus/sonnet/haiku aliases resolve to) and a small/fast model for background work like title generation (ANTHROPIC_DEFAULT_HAIKU_MODEL). Set both so neither tier falls back to a model your Otari instance does not serve. For example, to run the CLI entirely on managed open-weight models (a capable main model and a smaller background model):

export ANTHROPIC_BASE_URL="https://api.otari.ai"
export ANTHROPIC_AUTH_TOKEN="tk_your_otari_token"
export ANTHROPIC_DEFAULT_OPUS_MODEL="mzai:moonshotai/Kimi-K2.6"
export ANTHROPIC_DEFAULT_SONNET_MODEL="mzai:moonshotai/Kimi-K2.6"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="mzai:Qwen/Qwen3-32B"
claude

Caveats with non-Claude models

Routing works at the protocol level, but quality does not transfer automatically. Claude Code's prompts and tool-use loop are tuned for Claude models; other models run through the same Otari path but may produce weaker tool calling and agentic behavior. Anthropic-only features (extended thinking, prompt caching via cache_control) have no equivalent on most providers and are dropped in translation. Treat non-Claude models as usable, not equivalent.

What Otari provides

• POST /v1/messages — the Anthropic Messages endpoint Claude Code drives, with streaming and tool use.
• POST /v1/messages/count_tokens — Claude Code calls this every turn to keep the prompt within the context window. Otari counts locally (no provider call, no budget debit) and returns {"input_tokens": N}. The count is an approximation; it is used only to gauge headroom, not for billing.

Token usage and cost from /v1/messages are recorded and reconciled exactly as for /v1/chat/completions, so Claude Code sessions show up in usage and budgets like any other client.