Vane

What Vane does

Vane is an answer engine you run yourself. It was called Perplexica until the repository was renamed; the old GitHub address now redirects to the new one. You type a question, Vane decides whether it needs to search, queries the web through SearXNG, has a language model read the results and writes an answer with citations you can click through. Speed mode returns a quick answer, while Quality mode runs a longer research pass. You can steer it towards discussions or academic papers, restrict it to particular sites, ask about files you upload, and see images and videos next to the text. Small widgets show the weather, stock prices or a calculation when they match the question.

It is built for search-style questions, not as a general assistant: there is no tool or MCP support, no code execution and no image generation.

Models and providers

Vane uses separate chat and embedding models. For local models it connects to Ollama, LM Studio or a Lemonade server, and it can compute embeddings inside the app with Transformers.js. Hosted options are OpenAI, Anthropic, Google Gemini and Groq, and the OpenAI provider takes a custom base URL, so any OpenAI-compatible server such as LocalAI or llama.cpp works too. With a local model, the search queries still go to SearXNG and on to the engines it uses, but the question and the pages read never reach an AI company.

Privacy and data

Chats, settings and search history are stored in a SQLite database in the app’s data folder, which the Docker setup keeps in a named volume. No telemetry or analytics code turned up in the source, and none is documented. The project chose SearXNG so that searches reach several engines without identifying you; a commercial search API is not required.

Self-hosting Vane

The default Docker image bundles SearXNG, so one docker run command with a data volume starts everything on port 3000. API keys, models and the search backend are then set in a setup screen. A slim image is available if you already run SearXNG; your instance must have JSON output and the Wolfram Alpha engine enabled. You can also build from source with npm. When Vane runs in Docker and Ollama runs on the host, Vane must reach it at host.docker.internal (Windows and macOS) or the host’s private IP (Linux), and Ollama must listen on all interfaces.

Vane has no user accounts or login; authentication is listed as an upcoming feature. Anyone who can reach the port can use it and your API keys, so put it behind a reverse proxy with authentication before exposing it beyond your own network.

Hosted version

There is no official hosted service. The README links one-click templates on third-party platforms such as Sealos, RepoCloud, ClawCloud and Hostinger.

Limitations

Answer quality depends heavily on what SearXNG returns and on the model: small local models summarize less reliably than large hosted ones. It is single-user with no access control. Some guides and deployment templates still use the Perplexica name, which can confuse searches for help.

Who it suits

Vane suits people who want Perplexity-style cited answers on their own server, especially with a local model, and who can run Docker. For a full assistant that also searches the web, see LibreChat or Khoj; the Perplexity alternatives page compares the options.