Deployment

Use this guide after the Quickstart. The quickstart gets a standalone Otari running locally and walks through the first authenticated request. This page picks up from there with deployment-specific setup: hybrid mode, optional services, and environment-based configuration.

Otari is distributed as a Docker image on Docker Hub.

Prerequisites

• Docker and Docker Compose

Standalone deployment notes

For the local standalone path, start with the Quickstart.

When turning that setup into a longer-lived deployment:

• Set database_url to a durable Postgres instance.
• Keep master_key for management endpoints, but use generated API keys for application traffic.
• Add pricing entries for every model you want budget enforcement on.
• Point container health checks at /health and /health/readiness.

Deploy on Railway

For a hosted standalone deployment without local setup, use the one-click Railway template. It stands up two services: Otari (docker.io/mzdotai/otari:latest, target port 8000, healthcheck /health) and a managed Postgres, wired together with OTARI_DATABASE_URL=${{Postgres.DATABASE_URL}}.

You set at least one provider key (the form prompts for OPENAI_API_KEY; add a variable like ANTHROPIC_API_KEY, MISTRAL_API_KEY, or GEMINI_API_KEY to use another provider). The master key is auto-generated and OTARI_REQUIRE_PRICING=false is pre-set so an env-only deploy is usable out of the box. The template definition lives in deploy/railway/.

Connect to otari.ai

In hybrid mode, Otari delegates provider routing, authentication, and usage tracking to otari.ai. No local database or provider credentials are needed.

1. Create a minimal config file

host: "0.0.0.0"
port: 8000

No providers block, no database_url, no master_key.

2. Set your otari.ai credentials

You need the gateway token (gw-...) for this Otari instance from otari.ai. In otari.ai, go to Organisation > Gateways, create or open a gateway, then click Create token. This is not the per-request user token (tk_...) that clients send in Authorization: Bearer ....

Pass it as an environment variable. Create a .env file:

OTARI_AI_TOKEN=gw_your_token_here

3. Start Otari

docker run --rm \
  -p 8000:8000 \
  --env-file .env \
  -v "$(pwd)/config.yml:/app/config.yml:ro" \
  mzdotai/otari:latest \
  otari serve --config /app/config.yml

No postgres container is needed -- otari.ai handles storage.

4. Verify

curl http://localhost:8000/health

The response includes platform reachability status:

{"status": "healthy", "mode": "hybrid", "platform_reachable": "yes"}

Check readiness:

curl http://localhost:8000/health/readiness

Then verify a chat request using an otari.ai user token:

curl http://localhost:8000/v1/chat/completions \
  -H "Authorization: Bearer <your-otari-user-token>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o",
    "messages": [{"role": "user", "content": "Say hello in one short sentence."}]
  }'

Optional services

Otari supports two opt-in services via Docker Compose profiles.

Code execution sandbox

A sandboxed Python REPL for otari_code_execution tool calls:

docker compose --profile code-exec up -d

Web search

A SearXNG-based web search backend for otari_web_search tool calls:

docker compose --profile web-search up -d

SearXNG's free engines (DuckDuckGo, mojeek, qwant, …) rate-limit/CAPTCHA automated queries by IP, so they can be unreliable. For dependable results, use a licensed search API instead. A ready-to-run Brave Search adapter ships in scripts/web-search-brave-adapter/:

export BRAVE_API_KEY=...   # from https://brave.com/search/api/
export OTARI_WEB_SEARCH_URL=http://brave-adapter:8080
docker compose --profile web-search-brave up -d --build brave-adapter otari

A ready-to-run Tavily adapter also ships in scripts/web-search-tavily-adapter/:

export TAVILY_API_KEY=...   # from https://tavily.com/
export OTARI_WEB_SEARCH_URL=http://tavily-adapter:8080
docker compose --profile web-search-tavily up -d --build tavily-adapter otari

WebSearchBackend is URL-configured, so any service exposing a SearXNG-compatible /search?format=json endpoint works — copy the adapter to front Exa, Serper, etc.

Both code-exec and web-search profiles can be combined:

docker compose --profile code-exec --profile web-search up -d

When a profile is not running, Otari returns a 502 to requests that try to use that tool.

Environment variables

Provider API keys can be passed as environment variables instead of putting them in config.yml:

Variable	Description
OPENAI_API_KEY	OpenAI API key
ANTHROPIC_API_KEY	Anthropic API key
MISTRAL_API_KEY	Mistral API key
GEMINI_API_KEY	Google Gemini API key
OTARI_PORT	Otari server bind port (default: 8000)
OTARI_MASTER_KEY	Master key for management endpoints
OTARI_AI_TOKEN	Platform token from otari.ai

See Configuration for the full reference.