Vane Open-Source Perplexity Alternative: How to Run It (2026)

Vane (formerly Perplexica) is an open-source, self-hosted answering engine designed as a private alternative to Perplexity AI. It pairs a local SearXNG meta-search engine with local models via Ollama or commercial models via API keys, returning web-grounded answers with citations without monthly subscription fees. You can deploy it locally using Docker Compose in about five minutes, though running capable models requires dedicated hardware and ongoing maintenance of search engine scraping endpoints.

Why Vane is gaining traction right now

Developer interest in self-hosted search has surged as subscription costs and privacy concerns mount. Perplexity Pro costs $20 per month (or $200 per year as of November 2026), and queries run entirely through proprietary cloud servers. For developers, researchers, and privacy-conscious users handling internal documentation or proprietary data, sending every query to a commercial platform is a compliance non-starter.

The project created by developer ItzCrazyKns was originally called Perplexica, racking up more than 18,000 GitHub stars before rebranding to Vane. The rename reflects its expanded scope: serving as a customizable navigation instrument through web data rather than just a clone of an existing product. It allows users to bring their own models, customize search backends, and maintain full control over query history.

How to set up and run Vane with Docker

Setting up Vane requires a machine with Git, Docker, and Docker Compose installed. You also need an inference provider: either a local Ollama instance running on your machine or API keys from providers like Anthropic, OpenAI, or Google.

Step 1: Clone the repository

Open your terminal and clone the official repository:

git clone https://github.com/ItzCrazyKns/Vane.git
cd Vane

If you previously cloned ItzCrazyKns/Perplexica, updating your remote URL to the new Vane repository points your build to the latest releases.

Step 2: Configure your environment

Copy the sample configuration file into a live configuration:

cp sample.config.toml config.toml

Open config.toml in your code editor. This file controls the web search backend, model provider, and port mappings.

If you are using Ollama, ensure your Ollama endpoint points to your host machine:

[GENERAL]
PORT = 3000
SIMILARITY_MEASURE = "cosine"

[MODELS.OLLAMA]
API_URL = "http://host.docker.internal:11434"

If you are using commercial API keys instead of local weights, enter your credentials in the corresponding provider blocks:

[MODELS.OPENAI]
API_KEY = "sk-your-openai-api-key"

[MODELS.ANTHROPIC]
API_KEY = "sk-ant-your-anthropic-api-key"

Step 3: Start the Docker stack

Vane bundles the web user interface, backend application, and a SearXNG meta-search instance. Start all services in the background:

docker compose up -d

Docker downloads the necessary container images, configures SearXNG, and starts the Vane server. You can monitor startup logs by running docker compose logs -f.

Step 4: Access the interface and pick a focus mode

Once the containers finish initialising, navigate to http://localhost:3000 in your browser.

Vane provides specialised focus modes that alter how SearXNG gathers information and how the language model formats responses:

• Web Search: General web search using engines configured in SearXNG.
• Academic Search: Targets scientific papers and academic preprint servers.
• Writing Assistant: Disables web retrieval to generate text directly from the model.
• YouTube Search: Searches video transcripts to extract specific moments and quotes.
• Reddit Search: Scrapes public community discussions for user opinions.

The hidden friction points of self-hosting Vane

While running a private search engine sounds ideal, self-hosting Vane involves practical trade-offs that often surprise users moving over from hosted tools.

1. SearXNG rate limits and IP blocks

Vane relies on SearXNG to fetch live results by scraping major search engines like Google, Bing, and DuckDuckGo. Public search engines actively detect and block requests coming from datacenter IP addresses or frequent residential queries.

Without configuring rotating residential proxies, SearXNG instances frequently encounter CAPTCHAs, bot challenges, or silent rate limits. When SearXNG fails to retrieve web snippets, Vane falls back to hallucinating answers without live citations.

2. Hardware constraints for local models

Running Vane completely offline requires hosting a local language model via Ollama. Small models like Llama 3.1 8B or Qwen 2.5 7B fit comfortably on consumer laptops with 8 GB to 16 GB of RAM, but they often struggle to synthesize information from multiple long web pages accurately. They frequently drop citation markers or misattribute facts.

Frontier-grade reasoning models that can match Perplexity Pro, such as 70B parameter models, demand at least 48 GB of unified memory on Apple Silicon or high-end dedicated GPUs.

3. Pay-per-token API bills accumulate quickly

Many users resolve the local hardware problem by plugging Anthropic or OpenAI API keys into Vane. However, search engines feed extensive context into each request.

Every search query pulls in several web pages of scraped text. In a multi-turn conversation, re-sending that full search context on each follow-up message can burn through thousands of input tokens per exchange. A researcher running 20 to 30 deep searches a day on commercial API keys can easily spend $25 to $50 per month, eliminating the cost advantage over a flat-rate subscription.

4. Single-pass retrieval vs multi-step deep research

Vane executes a single-pass search: it takes your query, runs one set of search terms through SearXNG, parses the top results, and generates a response.

It does not perform autonomous deep research. If your research question requires exploring multiple branches, cross-referencing conflicting claims across 20 or more pages, or digging into nested links over several minutes, Vane stops after the initial search results.

When to use Zemith for multi-model research

For researchers, founders, and professionals who need deep web search with frontier models but do not want to maintain Docker containers, debug SearXNG proxies, or manage separate API billing, Zemith provides a zero-setup alternative.

Zemith combines frontier model choice with autonomous research and document export:

• Flagship models included: Switch between Claude 5.5 Sonnet, GPT 6.1 Sol, Gemini 4 Argon, and Perplexity Sonar within the same conversation on the Plus plan ($15.99 per month or $179.88 per year).
• Autonomous Deep Research: Instead of a single search pass, Zemith's Deep Research agent formulates an explicit investigation plan, requests your approval, and reads about 24 sources over up to 10 minutes in the background. It continues running even if you close your browser or phone app, delivering an exhaustive report with full footnotes for 100,000 credits per run.
• Export to finished Office files: Rather than copying raw markdown tables, you can generate and download native Word documents (.docx), Excel models (.xlsx), or presentation slides (.pptx) directly from your research chat.
• Library integration: Save research papers, transcripts, and web sources into a Library Project holding up to 50 sources, allowing you to cross-examine complex materials with full active-recall citations.

Self-hosted Vane vs Perplexity Pro vs Zemith

The right choice depends on whether your priority is zero-trust network isolation, a polished turnkey search engine, or multi-model research that produces finished files.

Why did Perplexica rename to Vane?

The developer, ItzCrazyKns, renamed Perplexica to Vane to establish a distinct open-source identity rather than being perceived indefinitely as a secondary clone of Perplexity AI. A wind vane points direction through moving air; the name reflects an instrument that directs users toward accurate answers amid noisy web data.

The underlying repository at github.com/ItzCrazyKns/Vane maintains the same codebase, license, and Docker deployment structure, with active ongoing updates to its focus modes and search extractors.

What hardware do you need to run Vane with Ollama?

To run Vane effectively on local hardware without cloud API keys, your system requirements depend on the model size you choose:

• Minimum (8B models like Llama 3.1 8B): An 8-core CPU, 16 GB of system RAM (or 8 GB of VRAM on an NVIDIA GPU), and 20 GB of free storage. Answers will be fast, but complex multi-source synthesis will frequently miss details.
• Recommended (14B to 32B models like Qwen 2.5 32B): Apple Silicon with 32 GB of unified memory, or a dedicated GPU with at least 16 GB of VRAM.
• Frontier parity (70B models): Apple Silicon with 64 GB+ unified memory, or dual GPUs with 48 GB total VRAM.

If your machine does not meet these specifications, running local models will result in sluggish token generation and memory exhaustion errors during search summarisation.

How does Vane compare to Perplexity Pro for research?

Perplexity Pro remains significantly more capable for complex exploratory research than a standard Vane installation. Perplexity maintains its own web-scale index and scraping infrastructure, bypassing the CAPTCHAs and bot detections that frequently degrade SearXNG instances.

Furthermore, Perplexity Pro includes automated deep research that recursively queries the web to clarify ambiguous answers. Vane executes a single prompt-and-search cycle. Unless you write custom orchestrations, Vane cannot autonomously plan, branch, and compile multi-page dossiers across dozens of websites.

For researchers evaluating alternatives to Perplexity Pro, our breakdown of Perplexity vs ChatGPT for research analyzes how citation accuracy and multi-step reasoning compare across major platforms.

When should you self-host Vane instead of using a hosted tool?

Self-hosting Vane is the best approach if you fall into one of two categories:

1. Strict air-gapped or confidential environments: If you work with patient records, legal evidence, or sensitive source code that policy forbids sending across any third-party network, Vane paired with a local Ollama instance guarantees that zero packets leave your local network.
2. Homelab hobbyists: If you enjoy maintaining self-hosted services, configuring SearXNG instances, tuning proxy pools, and managing Docker containers, Vane is an engaging and capable open-source project.

However, if your primary goal is producing thorough, cited research quickly without troubleshooting Docker network bridges, dealing with blocked search queries, or watching un-cached API tokens drain your balance, a managed multi-model service is far more productive.

The bottom line

Vane is an impressive open-source achievement that proves self-hosted, cited AI search is possible on consumer hardware. For privacy purists and homelab developers, it provides a solid foundation for private answering engines.

For everyday academic, market, and technical research, the ongoing maintenance of SearXNG scraping endpoints and the steep hardware requirements of local frontier reasoning make self-hosting a heavy commitment.

If you want access to frontier models like Claude 5.5 Sonnet and Gemini 4 Argon alongside automated Deep Research and native Office document downloads, Zemith Plus delivers a complete research environment for $15.99 per month without the infrastructure burden. Explore how Zemith stacks up against standalone search engines on our Zemith vs Perplexity comparison.