This repo is the follow-through. A search API built on the QVAC SDK, where Brave results get cleaned by your own local QVAC LLM — never a cloud server — so agents running on Tether's edge stack can read the live web without breaking the "data never leaves your hardware" principle.
We call it the open-web hop for QVAC agents.
✅ v0.2.2 live at qsearch.pro (2026-04-19). 4 endpoints:
/search,/news,/context,/health+ MCP tool. The live demo cleans on the demo server so you can try it without installing. Self-hosted instances clean on your own machine — that's the design. Daily log: @TheTieTieTies · Roadmap: ROADMAP.md · PRD: docs/PRD.md
# 1. Clone
git clone https://github.com/theYahia/qsearch.git
cd qsearch
# 2. Get a Brave Search API key ($5/month, ~1000 queries)
# → https://brave.com/search/api/ → sign up → copy key
# 3. Create .env.local and paste your key
cp .env.example .env.local
# Open .env.local and replace "your_brave_data_for_ai_key_here" with your actual key
# 4. Install & run
npm install # first run downloads Qwen3-0.6B (~364MB, cached after)
npm start # → qsearch v0.2.2 listening on http://localhost:8080
# 5. Test
curl -X POST http://localhost:8080/search \
-H "Content-Type: application/json" \
-d '{"query": "qvac sdk", "n_results": 2}'Brave API key is BYOK — you get your own key, it stays in your .env.local, never leaves your machine. qsearch doesn't proxy, relay, or store your key.
Tether's edge-first open-source stack came together over the past week:
- QVAC SDK (2026-04-09) — local LLM inference on phones, laptops, Raspberry Pi
- WDK (2026-04-13) — self-custodial wallet toolkit
- QVAC Workbench — local-document Q&A desktop app
What's missing is the open-web hop. An agent running on QVAC can answer from its own files, but the moment it needs to read the live web, it either (a) calls Exa/Tavily/Sonar — which means sending the query and seeing the cleaned result through a cloud server — or (b) parses raw HTML by hand.
qsearch is the primitive that closes that gap on the user's own hardware.
flowchart LR
A[Your agent] -->|query| Q[qsearch]
Q -->|1 fetch| B[Brave Search API]
B -->|raw results| Q
Q -->|2 clean locally| L["@qvac/sdk<br/>local LLM"]
L -->|cleaned markdown| Q
Q -->|structured JSON| A
style L fill:#86efac,stroke:#16a34a,color:#000
style Q fill:#93c5fd,stroke:#2563eb,color:#000
The green node is the whole story. The LLM cleaning step — the part that reads the page, extracts the answer, decides what matters — runs on the user's device, not on our server. It's architectural, not a privacy-policy promise. You can verify it by reading the code.
Not because it's trendy. Because it's the only search backend where the whole architecture holds:
- Independent index. Brave crawls its own web — not a Google or Bing wrapper. qsearch is a real alternative to the big-cloud APIs, not a thin reskin.
- Data-for-AI tier, BYOK. Brave's commercial tier explicitly supports AI transformation of results, removing the ToS grey zone that blocks agent apps on other providers.
- No query profiling upstream. Brave's business model doesn't depend on tracking queries. The data-hygiene story is consistent end-to-end: Brave doesn't track, qsearch doesn't clean in the cloud, the agent stays local.
- Not owned by a cloud giant. Using Google/Bing to power a Tether-aligned, edge-first primitive would be architecturally incoherent. Brave is independent — that matches the ethos of the stack we're building on.
- Stable API, good docs. Less time fighting the backend, more time on the cleaning layer — which is where the differentiation lives.
We're not locked to Brave forever — v2 may add SearXNG or Mullvad Leta as drop-in providers. But for the MVP, one backend that fits the thesis end-to-end > three backends that fight it.
| Exa | Tavily | Sonar | Brave API | SearXNG | qsearch | |
|---|---|---|---|---|---|---|
| OSS core | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
| LLM cleaning | ✅ (cloud) | ✅ (cloud) | ✅ (cloud) | ❌ | ❌ | ✅ (local) |
| Agent-first JSON | ≈ | ≈ | ≈ | ❌ | ❌ | ✅ |
| Self-hostable | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
| QVAC-native | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| BYOK upstream | ❌ | ❌ | ❌ | N/A | ✅ | ✅ |
qsearch is the first row where all of these are checked. That's the wedge — not better snippets, not faster ranking. Local cleaning on OSS, as a primitive for agents. The intersection didn't exist until now.
| Endpoint | What | Brave source |
|---|---|---|
POST /search |
Web search + QVAC cleaning | /web/search (1-20 results) |
POST /news |
News search + cleaning | /news/search (1-50 results) |
POST /context |
Deep page extraction + cleaning | /llm/context (2-28 snippets/source) |
GET /health |
Server status | — |
| Parameter | Type | Description |
|---|---|---|
query |
string | Search query (required) |
n_results |
number | Results count (default: 3) |
freshness |
string | pd (day), pw (week), pm (month), py (year), or YYYY-MM-DDtoYYYY-MM-DD |
search_lang |
string | Language: "en", "ru", etc. |
country |
string | Country: "us", "ru", etc. |
curl -X POST http://localhost:8080/search \
-H "Content-Type: application/json" \
-d '{"query": "qvac sdk", "n_results": 2}'{
"query": "qvac sdk",
"model": "QWEN3_600M_INST_Q4",
"brave_ms": 819,
"results": [
{
"url": "https://qvac.tether.io/",
"title": "QVAC - Decentralized, Local AI in a Single API",
"description": "QVAC is Tether's answer to centralized AI...",
"cleaned_markdown": "QVAC is a decentralized, local AI platform built on Tether, offering a new paradigm where intelligence runs privately, locally, and without permission on any device.",
"clean_ms": 1420
},
{
"url": "https://tether.io/news/tether-launches-qvac-sdk...",
"title": "Tether Launches QVAC SDK...",
"description": "QVAC SDK is a unified software development kit...",
"cleaned_markdown": "Tether.io launched the QVAC SDK, a unified AI development kit enabling AI training and evolution across any device and platform.",
"clean_ms": 1008
}
]
}brave_ms = Brave fetch latency. clean_ms = local LLM inference per result. The green node in the diagram above — that's where clean_ms happens, on your machine.
Stack:
- Runtime: Node.js ≥20 (
@qvac/sdkbundles bare worker — no system install needed) - Backend: Brave Search API, BYOK (
BRAVE_API_KEYin.env.local) - LLM:
@qvac/sdkwith Qwen3-0.6B Q4 (~364MB, downloads once, cached locally) - License: Apache-2.0
- Cold start. Loading a local LLM takes seconds. qsearch is best run as a long-lived local daemon, not a cold-fired lambda.
- Single provider in v1. Brave only. More providers are v2.
- Self-host first. qsearch.pro is a public demo, but the design assumes you run your own instance. If you want zero-ops, Exa and Tavily exist and are good.
- The wedge is architecture, not ranking. qsearch won't out-rank Exa on snippet quality. It wins when you care that cleaning runs on your hardware, not theirs.
A new commit, demo, or writeup ships every day until 2026-04-21.
- 🌐 Live demo: qsearch.pro — try it in your browser
- ⭐ Star this repo — v0.2.2 live with 4 endpoints + MCP tool
- 🐦 X thread: @TheTieTieTies — daily updates
- 🗺️ Full 7-day plan: ROADMAP.md
- 📝 Feature requests for v2: open an issue
Apache-2.0 — same as QVAC itself. See LICENSE.