Onyx: Self-Hosted AI Chat and Search Over Your Team's Apps (GitHub, Scanned)

Onyx is a self-hosted AI assistant that knows what is in your company's tools. It connects to more than 50 sources such as Google Drive, Slack, Confluence, Jira, SharePoint and GitHub, indexes them along with their permissions, and answers questions with citations, so people only get answers drawn from documents they could open anyway. On top of search it adds deep research, custom agents, web search, MCP actions, a code sandbox and voice.

It runs as Docker containers or on Kubernetes, works with hosted models (OpenAI, Anthropic, Gemini) or self-hosted ones (Ollama, vLLM, LiteLLM), and can be fully air-gapped. A Lite mode under 1 GB of memory gives you the chat interface and agents without document indexing. People reach it through web and desktop apps, a Slack or Discord bot, an MCP server for coding agents, and a Chrome extension.

The Community Edition is MIT-licensed; folders named ee hold the Enterprise Edition features (SSO, RBAC, analytics and similar) under a commercial licence. About 33,000 stars and 210 contributors, at v5.0.0. The full stack is heavy, with a search index, background workers and model servers, so plan a proper server for it.

• Repository: github.com/onyx-dot-app/onyx
• Licence: custom (Other)
• Language: Python. Stars: 33K. Forks: 4,646. Last push: Oct 26, 2026.
• Scan: safe, Oct 26, 2026, commit 616a4b1

Who it is for

Teams that want a ChatGPT-style assistant grounded in their own documents while keeping the data on their own infrastructure, and IT staff weighing self-hosted alternatives to enterprise AI search products.

Getting started

1. Install the Onyx CLI, which then walks you through a Docker deployment

curl -fsSL https://onyx.app/install_onyx.sh | bash

2. Or deploy the lightweight Lite mode without prompts

curl -fsSL https://raw.githubusercontent.com/onyx-dot-app/onyx/main/deployment/docker_compose/install.sh | bash -s -- --lite --no-prompt

The installer in the repository (deployment/docker_compose/install.sh) is only a bootstrap: it installs onyx-cli to ~/.local/bin (or /usr/local/bin as root), adds that folder to your PATH in your shell profile, and hands over to the CLI. Set DISABLE_TELEMETRY=true on the services if you do not want usage reporting.

Safety scan

We cloned onyx-dot-app/onyx at commit 616a4b1 on Oct 26, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• The 40 secret hits are test fixtures (fake Slack tokens and private-key markers) and the same placeholder text, BEGIN PRIVATE KEY, in nine translation files, where it is the hint shown in an SSO key field. The 14 bare-IP URLs are SSRF tests checking that Onyx refuses them.
• Fleet telemetry is on by default. Per the project's deployment/fleet-telemetry.md, each deployment reports to telemetry.onyx.app every few minutes under a random deployment key, including query counts and timings, connector settings and counts, and signup email domains. Setting DISABLE_TELEMETRY=true on all services stops it.
• The six pattern hits are the curl | bash installer and comments about it. The script downloads onyx-cli from the project's GitHub releases, writes a PATH line to your shell profile, and only suggests sudo if it cannot write to the chosen folder.
• 101 advisories, none critical: the Python backend's uv.lock (55, 21 high, in aiohttp, fsspec and similar), the Tauri desktop app's Cargo.lock, and three Go modules for the CLI and tooling.
• 52 workflows. One uses pull_request_target, an ownership check for the ee folders, and it does not check out the pull request. All 272 third-party actions are pinned to a commit; Dependabot, CodeQL and a security policy are present.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Python 987.1K, TypeScript 449.3K, JSON 266.5K, Go 104.3K, Markdown 76.6K, YAML 50.2K.

Questions

Is Onyx free?

The Community Edition is free and MIT-licensed, and covers search, chat, agents and actions. The Enterprise Edition adds features such as SSO, role-based access and analytics under a paid licence, and Onyx Cloud is a hosted version if you would rather not run it yourself.

Can Onyx run with a local model?

Yes. It supports self-hosted models through Ollama, vLLM and LiteLLM as well as the big hosted providers, and the embedding model can run locally too, which is what makes a fully air-gapped deployment possible.

What does Onyx need to run?

Docker or Kubernetes. The standard deployment runs several containers, including a search index, background workers and model servers, so give it a real server. Lite mode needs under 1 GB of memory but skips document indexing, so it is a chat interface rather than company search.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.