An open-source LLM control plane built by Mozilla.ai, released on July 6, 2026. The more AI providers you adopt, the more SDKs, API keys, and billing accounts sprawl across your stack, leaving your application buried in glue code that exists only to wire things together. Otari sits in front of all of it, consolidating 40+ providers behind a single OpenAI-compatible endpoint. On top of that, it handles per-workspace and per-key budget caps, centralized key management, usage logging, and automatic failover when a provider goes down. It is available both hosted and self-hosted — with self-hosting, your API keys and usage data never leave your own environment. Licensed under Apache-2.0.
Key Features
- One endpoint, 40+ providers: Point your application at a single OpenAI-compatible (and Anthropic-compatible) endpoint and swap the provider behind it. Being able to change models without rewriting application code is the core benefit
- Virtual API keys: Instead of handing out real provider keys, you distribute virtual keys issued by Otari to teams and applications. Each can be revoked individually, limiting the blast radius when someone leaves, a key leaks, or a proof of concept ends
- Budget caps enforced before the request runs: Limits can be set per user and per key, and requests that would exceed them are stopped before execution. No more discovering the overspend when the invoice arrives
- Usage and spend logs: The
/v1/usageendpoint and the admin dashboard show who used which model and how much. Useful material for audits and internal cost allocation - Guardrails and sandboxed tools: Prompt injection detection can be applied at the request level, and the gateway ships with sandboxed Python execution and web search as built-in capabilities
- Agent support (Agent Harnesses) and automatic failover: Built-in mechanisms designed for agent workloads, plus automatic switching to an alternate provider when one fails
- Hosted and self-hosted: Run it yourself with plain Docker, Docker Compose (with PostgreSQL), Render, or Railway. A hybrid setup that self-hosts while connecting to the otari.ai platform is also available
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Self-hosted | $0 (Apache-2.0) | The software itself is free. You bring your own provider API keys and pay each provider directly |
| Hosted (otari.ai) | Contact for details | When using providers supplied by Otari, billing is described as per-token charges drawn against your wallet |
Pricing is as of August 2026. Public information about the hosted plan’s billing model is limited, so check the official site and the documentation for current and accurate terms.
Pros & Cons
✅ Pros
- The OpenAI-compatible endpoint means most existing code and libraries carry over unchanged
- Apache-2.0 open source, so self-hosting keeps vendor lock-in low
- API keys and usage data can stay inside your own network, making it viable in environments with strict internal or client-side requirements
- Budget caps work as a pre-request block rather than after-the-fact analysis, structurally preventing cost accidents
- Virtual keys can be issued and revoked individually, which simplifies access control per team or project
⚠️ Cons
- Self-hosting means you take on the operational burden of containers, a database, and updates
- Inserting a gateway creates a potential single point of failure; designing for redundancy is on you
- Released in July 2026, so the surrounding ecosystem and body of real-world case studies are still young
- Little public information about hosted pricing, which often means you have to reach out before you can compare options
- Abstracting across providers can get in the way when you want to use a specific vendor’s newest features directly
Comparison with Similar Services
| Criteria | Otari | LiteLLM | OpenRouter | Cloudflare AI Gateway |
|---|---|---|---|---|
| Provider | Mozilla.ai | BerriAI | OpenRouter | Cloudflare |
| Form | OSS + hosted | OSS + paid tier | Hosted service | Hosted service |
| Self-hosting | Yes (Apache-2.0) | Yes | No | No |
| Main focus | Centralized keys, budgets, auditing | Unified SDK / proxy | Instant access to many models | Observability, caching, rate control |
| Budget controls | Pre-request block per user/key | Per-key limits | Credit balance based | Mainly rate limiting |
| Bring your own keys | Assumed (BYOK) | Assumed (BYOK) | Optional | Assumed (BYOK) |
Who Is It For
- Teams running several LLM providers side by side, where SDK and API key management has become unwieldy
- Organizations that want explicit AI spending caps per department or per project
- Environments with strict security requirements that cannot hand API keys and usage logs to an external SaaS
- Developers who want to avoid lock-in to one provider and keep room to switch models
- Development teams putting agents into production who now need failover and cost control against runaway loops
Summary
Otari consolidates the keys, billing, logs, and switching logic that tend to scatter across LLM integrations into a single gateway, as open source. If you use only one provider today, the benefit is thin — it starts to pay off once providers multiply, users multiply, and costs stop being predictable. A practical starting point is to bring up the self-hosted version with Docker, repoint an existing application’s endpoint at it, and observe the behavior. Since hosted billing terms are not widely published, confirm them with the vendor before a full rollout.