Vane

Where does your data go?

Nowhere you didn't put it — with one deliberate exception:

• Self-hosted by design. Vane runs as a single container on your hardware — the image bundles the app, the API, and its own private SearxNG search engine. No account, no telemetry requirement, search history stored locally.
• The unavoidable exception: the search itself touches the web. SearxNG is a metasearch engine — it queries other engines on your behalf precisely so your identity doesn't go to Google directly. That's a privacy technique, not a guarantee; queries still transit your server to whichever upstream engines you enable.
• Your model is your data boundary. Point it at Ollama and everything stays local end-to-end (search metadata excepted). Point it at OpenAI or Anthropic and your queries go to those vendors under their rules — the same trade the ChatGPT and Claude reviews document. The tool doesn't enforce privacy; your model choice does.
• MIT-licensed open source — auditable end to end, no data-retention clause to parse because the data never reaches a vendor.

What happens when it's wrong?

Vane publishes no accuracy benchmarks — answer quality is your model's, and citation quality is SearxNG's. The documented failure surface is operational:

• It's an engine, not a product team. No support desk — GitHub issues (232 open), a Discord, and your own troubleshooting. The community is large (36k+ stars) but nobody owes you an answer.
• The known failure mode is configuration, not hallucination: the docs' longest section is Ollama connectivity errors — the classic being Docker containers unable to reach localhost. The fix (documented in the README) requires Ollama listening on all interfaces, which is the same unauthenticated-network-exposure posture flagged in the Ollama review. Two local tools, one shared security homework.
• The honest quality comparison: with a local 7–8B model, answers are decent but a visible notch below the hosted engines on reasoning; with a frontier cloud model, output quality matches Perplexity — because it's the same models. The engine isn't the intelligence; the model is.
• Citations: sourced and linked like Perplexity, but Vane publishes no citation-accuracy auditing, and its SearxNG layer can surface low-quality sources. The same rule applies from the hosted engines: the citation is a lead, not proof.

What's it actually bad at?

• Zero-setup it is not. Docker helps enormously (one image, bundled SearxNG), but a business owner still deploys a container, configures model providers in a setup wizard, and fixes connectivity errors. The README ships Ollama, Lemonade, and local-OpenAI troubleshooting sections — because enough people hit them.
• No official support. Issues and a Discord community; no SLA, no phone-a-human, no vendor on the hook.
• Feature gaps vs the hosted engines: no mobile apps, no team or workspace layer, no enterprise admin, no agent platform. It answers questions; it doesn't run your operations. Authentication is explicitly on the roadmap — meaning multi-user deployments currently have no access control at all. Exposing it beyond localhost today means anyone who can reach the port can use it.
• The search-quality ceiling: SearxNG metasearch is privacy-strong but weaker than Perplexity's purpose-built index for some queries; Tavily and Exa support is listed as coming soon.
• Renamed from Perplexica (March 2026) — the ecosystem's docs, tutorials, and muscle memory still say Perplexica; searching either name finds it.

What does it cost really?

The honest economics: a self-hosted Perplexity whose price is the model bill. With a local model, total cost of ownership is the machine. With GPT or Claude, you pay their API rates — at which point the question isn't cost, it's control. Deployment option: one-click images exist (Sealos, RepoCloud, ClawCloud, Hostinger) if you'd rather run it on a $5–10/month VPS than on your own hardware — still self-controlled, just not on-site.

What does it replace — and what do you still do?

It replaces the Perplexity relationship — cited web answers with search modes (Speed, Balanced, Quality), academic and discussion source filters, image and video search, file uploads, and domain-restricted searches — entirely under your control, with your browser's search bar pointed at it if you like. You still own: deployment, updates (pull-and-recreate the container), model selection and its costs, and network security. There's no vendor to hold accountable; that's the trade.

What the ecosystem offers

• One container, three engines: the image bundles the front end, the API, and its own private SearxNG — nothing else to deploy. A slim image exists if you already run SearxNG.
• Bring any model: Ollama and LM Studio for local, OpenAI, Anthropic, Gemini, or Groq for cloud — mix and match per search. Speed, Balanced, and Quality modes trade latency against depth and cost.
• An API: the same search engine your browser uses, callable from your own apps (a documented REST endpoint). For an owner with a developer, Vane becomes infrastructure.
• Browser integration: set it as your browser's default search engine — one URL and your address bar becomes a private answer engine.
• File understanding: upload PDFs, text files, and images and ask questions about them.
• In the works: Tavily and Exa search backends, custom agents, and authentication — the most requested feature, and the current gap that limits shared deployment.

If I stop paying, what do I lose?

• Nothing — there's nothing to cancel. MIT open source, no account, no vendor. The container and its data volume are yours.
• The lock-in reality is minimal: your search history is in a local volume; your workflows are browser shortcuts; the models are whatever you already pay for. Deleting the container loses nothing you can't recreate.
• The one real dependency: SearxNG's upstream engine behavior — if upstream engines change what they return, search quality shifts and you're waiting on the community. The same open-source trade as everywhere.

Who is this NOT for?

Anyone who wants zero-setup — the hosted engines exist precisely for that. Businesses without anyone willing to run a container and own patching. And anyone expecting support: this is open source with a big community, not a vendor with a phone number.