Big-AGI

What Big-AGI does

Big-AGI is a chat workspace aimed at people who use several AI models side by side, such as engineers, researchers and analysts. Its distinctive feature is Beam: one prompt goes to several models in parallel, you compare the replies, and a merge step combines the strongest parts into one answer. The developer presents this as a way to catch hallucinations and get better results than any single model.

Around that it has the features of a full assistant: personas with their own instructions, attachments, web search with citations on models that support it, web page loading, image generation and editing, voice calls with personas, text-to-speech from several vendors, diagrams, split windows and a multi-chat mode that sends one message to several conversations. The interface has a dedicated mobile layout.

Models and providers

Big-AGI connects to more than 20 providers, including OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock, Mistral, DeepSeek, Groq, OpenRouter, Perplexity, xAI and NVIDIA NIM. For local models it supports Ollama, LocalAI and LM Studio, and any OpenAI-compatible endpoint can be added, with models and capabilities detected automatically.

The app’s MIT license does not make the models open. When you pick a hosted model, the prompt goes to that provider. Only local servers keep conversations on your hardware.

Privacy and data

Big-AGI is client-first. API keys and settings sit in the browser’s local storage and chat history in IndexedDB; the server mainly relays requests to model APIs and stores nothing unless you use shared chat links. This means chats stay on one device and disappear if browser data is cleared, so the documentation recommends backups.

Telemetry depends on how you install it. The project’s documentation states that the official prebuilt Docker images report anonymous usage data (page views and feature-usage events, without prompts or conversation content) to a Google Analytics property. The ID is fixed at build time, so an environment variable on docker run cannot switch it off; building the image from source sends nothing. PostHog and Google Analytics can be enabled for your own builds, and Vercel Analytics turns on automatically on Vercel.

Self-hosting Big-AGI

Big-AGI Open runs with Docker (the ghcr.io/enricoros/big-agi image or a local build), Docker Compose, Vercel, Cloudflare or Kubernetes. Basic use needs no database. Postgres or MongoDB is only required for shareable chat links, and web browsing needs a Puppeteer-compatible service such as Browserless. API keys can be preset on the server so users do not need their own.

There is no built-in user management. Password protection with HTTP Basic Auth requires renaming a middleware file and rebuilding, and the docs otherwise suggest platform-level protection such as Cloudflare Access or Vercel authentication.

Hosted version

big-agi.com offers a free tier and a paid Pro tier; both use your own API keys, and Big-AGI does not charge for model usage. Pro adds Sync across devices, which stores conversations on the server with per-user isolation. The README describes some hosted improvements that are not presented as part of the open edition.

Limitations

Self-hosters who use the official images send anonymous usage data unless they build from source. There are no accounts, roles or SSO, no documented document chat or MCP support, and no sync between devices in the open edition. It is a tool for one person per browser rather than a team platform.

Who it suits

Big-AGI suits individual power users who pay for several model APIs and want to compare them in one fast interface, with data kept in their own browser. Teams that need logins should look at LibreChat, and people focused on local models at Jan. See the AI chat guide.