Fireplexity is an open-source AI answer engine. It searches the live web, scrapes what it finds with Firecrawl, and streams back a cited answer — alongside news, images, and ticker charts.
Sources · 4
Answer
Firecrawl renders each page in a real headless browser before extraction, so client-side routes and SPA content resolve the same way they would for a visitor1. The rendered DOM is then converted to clean markdown, stripping nav and boilerplate so only the substance reaches the model24
Images
Ask next
What it does
Every response is assembled from live sources, not a training snapshot — and every claim carries a link you can open.
Each inline marker points to the numbered source the model was given for that sentence. Hover one and the source card surfaces — favicon, title and link — so checking a claim takes a second, not a new tab. Markers aren’t verified automatically, so here’s how to check one.
Firecrawl returns ranked results and their full markdown together, so the model reads the most relevant passages of each page instead of guessing from search snippets.
Every query also pulls news results, with outlet and publish date kept visible, so you can judge how stale an answer is.
Visual queries return a thumbnail grid with a full-size lightbox and the page each image came from.
Mention a company and Fireplexity spots the symbol, then drops a live TradingView chart in beside the answer.
Each answer ends with up to five generated next questions, drawn from what the sources actually covered.
How it works
No vector store, no nightly crawl, no index to keep warm. The pipeline runs per query, start to finish, in one request.
01
One call to Firecrawl's search endpoint asks for web, news and image results together, so a single ask covers all three surfaces.
02
Each result is rendered in a real browser, then stripped of nav, ads, and boilerplate — leaving LLM-ready markdown instead of raw HTML.
03
An open model on Groq reads the passages and streams an answer token by token through the AI SDK, tagging each claim with the numbered source it cites.
Why run your own
Fireplexity is the whole thing — prompt, pipeline, and UI — in a repo you control.
The system prompt, the citation format, the follow-up logic — all of it is a file in app/api. Nothing about how answers are shaped is hidden from you.
Deploy it inside your own infrastructure and sensitive searches never pass through a vendor's logs or analytics.
The model, the search provider, the renderer, the theme. Each sits behind its own boundary, so replacing one doesn't mean rewriting the rest.
Two API keys and a Vercel deploy. No per-user pricing, no plan tier gating the features you already have the code for.
The stack
Everything here is something you already know how to debug. No bespoke framework, no custom runtime.
Self-host
$ git clone https://github.com/firecrawl/fireplexity.git $ cd fireplexity $ npm install # two keys, both free to start $ cp .env.example .env.local $ echo "FIRECRAWL_API_KEY=fc-..." >> .env.local $ echo "GROQ_API_KEY=gsk_..." >> .env.local # Groq retired the Kimi K2 model the repo ships with $ perl -pi -e 's#moonshotai/kimi-k2-instruct#openai/gpt-oss-120b#g' app/api/fireplexity/search/route.ts $ npm run dev → ready on http://localhost:3000
Fireplexity is MIT licensed and the entire pipeline is a few hundred lines. Read it, fork it, point it at your own sources.
We test each guide on open-source AI search, self-hosting and Perplexity alternatives before it goes up. Get a short note when the next one is out.