> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # Open-source LLM gateway Otari routes 40+ providers - URL: https://www.notatechguy.com/open-source-llm-gateway-otari-routes-40-providers/ - Published: 2026-10-20T04:43:38.000Z - Updated: 2026-10-20T04:43:38.000Z - Description: Otari, an open-source LLM gateway that routes to 40+ providers with virtual keys and budgets, hit GitHub Trending with 674 stars. - Author: Marcello Babbili - Tags: Technology & AI, AI Models, OpenAI Mozilla-ai pushed Otari onto GitHub's daily trending list on 19 October 2026, gaining 39 stars to reach 674 [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). The project is a self-hosted LLM gateway that speaks OpenAI's API protocol, exposing one endpoint that fans out to over 40 model providers \[S1, P2\] — a provider count that comes from the maintainers' own documentation and has not been independently verified. Most teams working with several LLM vendors juggle separate API keys for each, with no central view of spend or usage. Otari inserts itself between your application and the providers, issuing virtual API keys — sub-keys you generate that point to your real provider credentials, alongside per-key spending caps and usage logs in a single interface \[S1, P2\]. The maintainers say it is the same proxy server that runs otari.ai, now released for self-hosting [P²](https://github.com/mozilla-ai/otari/blob/main/docs/index.md?ref=notatechguy.com). **My read:** This is the first self-hosted LLM gateway I've seen that bundles virtual keys and budget enforcement into a package you run yourself, rather than locking those features behind a hosted service. The 40+ provider figure is the vendor's own count with no third-party confirmation, and 674 stars is early-stage territory. What I would watch is whether anyone outside the mozilla-ai org deploys this in production and reports back on stability. The gateway is written in Python and routes provider calls through the any-llm library [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). It runs in two modes: Standalone, where one process handles both management and the actual model calls using local storage, or Hosted, which links to otari.ai's infrastructure for organisations managing multiple teams [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). ### Docker, Railway, and Homebrew cover the install paths A Docker image is published as mzdotai/otari:latest, and the Compose setup runs PostgreSQL with optional profiles for a code sandbox, web search, and guardrail services (content filtering and safety checks) [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). There is a one-click deployment template for Railway, and a command-line tool installable via Homebrew [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). The project can operate standalone with a local database and your own provider credentials, or link to otari.ai's hosted infrastructure [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). The star count, the provider list, and the security posture are all maintainer self-description; no independent security audit has been disclosed, and no third-party deployment review has been published. The GitHub organisation is mozilla-ai, which is not the same as the Mozilla Foundation or Mozilla Corporation. ### Multi-vendor teams get a self-hosted alternative to hosted gateways A team building an AI feature that calls models from OpenAI, Anthropic, and a local open-weight deployment, and needs to cap spend per project or per customer, is the obvious first user. Today that team either writes custom routing code or pays a hosted gateway service. Otari offers a third path: run the proxy yourself, hand virtual keys to each developer or service, and set budget ceilings that block calls when the limit hits. For a platform engineer who currently maintains separate credential vaults per vendor, deploying Otari would replace that sprawl with one admin panel where virtual keys, per-project spend caps, and call logs sit together \[S1, P2\]. The open-weight ecosystem has been moving in this direction. OpenAI released gpt-oss-120b and gpt-oss-20b as open-weight models under Apache 2.0 [P⁴](http://github.com/openai/gpt-oss?ref=notatechguy.com), and projects like OpenChat have been pushing open-source model quality for years [P⁵](https://github.com/imoneoi/openchat?ref=notatechguy.com). As the number of usable models grows, the routing and cost-control layer becomes the bottleneck. The same day Otari trended, [Cline's open-source coding agent hit 70,000 GitHub stars](https://www.notatechguy.com/cline-open-source-coding-agent-hits-70-000-github-stars/), a reminder that developer tooling around mixed model stacks is where the activity is. For a team evaluating Otari today, the fastest path is the Docker image or the Railway template. The Homebrew CLI (`brew install mozilla-ai/tap/otari`) gets you a local instance for testing [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). As of 19 October 2026, the repo had 674 stars and no published independent reviews [S¹](https://github.com/mozilla-ai/otari?ref=notatechguy.com). --- *Sources: [S1 — mozilla-ai/otari: Open-source, OpenAI-compatible LLM gateway you run y](https://github.com/mozilla-ai/otari?ref=notatechguy.com) · [P2 — docs/index.md](https://github.com/mozilla-ai/otari/blob/main/docs/index.md?ref=notatechguy.com) · [P3 — Trending Papers - Hugging Face](https://huggingface.co/papers/trending?ref=notatechguy.com) · [P4 — openai/gpt-oss](http://github.com/openai/gpt-oss?ref=notatechguy.com) · [P5 — imoneoi/openchat](https://github.com/imoneoi/openchat?ref=notatechguy.com)* --- *Written from 5 sourced items, 4 of them primary.*