search_free
Search
Free engines first, merged by rank, three results a query, each with the engine that found it.
What is the Model Context Protocol (MCP)?modelcontextprotocol.io · via SearXNG (bing, yahoo, google cse)
Web access for AI agents.
Only the lines that answer.
Search many engines as one, read pages and PDFs, and get back the passages that answer the question, each with its source.
Loading a recorded run…
Anonymous demo: up to 12 runs a minute, 3 search results, crawls up to 4 pages (~20 s), public addresses only. No sign-up.
1,000 free credits · 8 lanes, 46 providers · MCP, Python SDK and HTTP · public beta
[ 02 / 09 ] · Clients
Works with the agents you already run.
- Claude Code
- Codex
- Cursor
- Claude Desktop
- Hermes
- any MCP client
MCP server · Python SDK · plain HTTP
[ 03 / 09 ] · Core
Three calls. Three kinds of answer.
scrape / fetch
Read
A page or a PDF comes back as markdown, with page markers and ruled tables. A scanned file is flagged as needing OCR instead of being guessed.
# asyncio — Asynchronous I/Odocs.python.org · fetched directly, no browser
fetch(url, query=…)
Ask a page
Pass a question and the page becomes the passages that answer it, in page order, each with its heading and offset.
asyncio is a library to write concurrent code using the async/await syntax.passage 1 · heading “asyncio” · docs.python.org
# your agent calls the tool
search_free(query="what is the Model Context Protocol")
fetch(url="https://docs.python.org/3/library/asyncio.html", query="what is asyncio for")
# pip install clawifi from clawifi import Clawifi client = Clawifi() # reads CLAWIFI_API_KEY hits = client.search_free("what is the Model Context Protocol") page = client.fetch("https://docs.python.org/3/library/asyncio.html", query="what is asyncio for")
curl https://api.clawifi.qzz.io/search/free -H "Authorization: Bearer $CLAWIFI_API_KEY" -H "content-type: application/json" -d '{"query": "what is the Model Context Protocol"}'
[ 04 / 09 ] · Trace
Every step on the record.
Each run says which engine was asked, which answered, what was merged and how long it took. When only one engine answers, it says so.
- Failed engines are shown, not hidden. A paused or rate-limited engine stays in the trace, dimmed, with the reason.
- Sources on every line. Every result carries its URL and the engine that found it.
- Nothing invented. If the live API can’t answer, the box replays a dated recording and labels it.
- asked SearXNG: “what is the Model Context Protocol”
- engines inside it: bing, yahoo, google cse
- 3 results, merged by rank
- sources: modelcontextprotocol.io, en.wikipedia.org, databricks.com
- answered end to end5.3 s
[ 05 / 09 ] · Measured
Measured, with the method attached.
38 of 40
searches landed on target
40 prospecting queries on the live playground, no paid key, 3 results each: all 40 answered, median 3.8 s. Measured 20 Oct 2026 from one location.
8 of 8 2 of 2
text PDFs read · scanned PDFs flagged
Every text PDF came back as markdown; both scanned files were flagged as needing OCR instead of returning fragments. Published 13 Oct 2026.
3,746919
tokens per page, median, with passages
The answer was in the top 3 passages on 26 of 29 pages. Published 13 Oct 2026.
[ 06 / 09 ] · Lanes
8 lanes, one API.
Each kind of question goes to the sources that hold it, free first.
| Lane | Providers | Examples |
|---|---|---|
| Free web search | 13 | SearXNG, DuckDuckGo, Marginalia, … |
| Sitemap, RSS, and docs discovery | 5 | robots.txt, Sitemaps, Page links, … |
| Open scholarly archives | 8 | arXiv, Crossref, Zenodo, … |
| Academic workflows | 5 | arXiv, PubMed, Semantic Scholar, … |
| Package registries | 8 | PyPI, npm, crates.io, … |
| Dataset registries | 4 | Hugging Face, OpenML, Dataverse, … |
| Code intelligence | 1 | GitHub |
| Domain and identity OSINT | 3 | Domain records, Username search, Email checks |
Also in the APIBrowserMonitorCrawlMapBatchSocial threadsDomain reconDatasetsPackages
Catalogue as of 2026-10-13 · every source →
[ 07 / 09 ] · Onboarding
Three steps to your agent’s first answer.
Which agent will use Clawifi?
Install, get a key, add it to your client:
pip install clawifi clawifi signup --email [email protected]
Restart the client and ask it something that needs the web:
“Use clawifi to search for the latest Python release notes and read the top result.”
[ 08 / 09 ] · Pricing
Start free. Pay as you grow.
Free
$0/ month
1,000 credits a month
Start freeStarter
$19/ month
10,000 credits a month
Choose StarterPro
$99/ month
125,000 credits a month
Choose ProScale
$299/ month
500,000 credits a month
Choose ScaleMost calls cost 1 credit · browser renders 3 · failed calls and cache hits are free · every price →
[ 09 / 09 ] · FAQ
Questions, answered plainly.
What does Clawifi return?
Search returns ranked results with the URL and the engine that found each one. Read returns a page or PDF as markdown. Read it with a question and you get only the passages that answer it, each with its heading and its place in the page.
Which agents can use it?
Any MCP client: Claude Code, Claude Desktop, Cursor, Codex, Hermes and others. There is also a Python SDK and a plain HTTP API.
Which engines does search use?
Free engines first. The hosted playground runs a private SearXNG that queries several web engines and merges them by rank; paid providers are only used when you enable them.
What happens when a page can’t be read?
It says so. A blocked page is reported as blocked, and a scanned PDF is flagged as needing OCR instead of being passed off as text.
How are credits counted?
Most calls cost 1 credit. A page that needs the browser costs 3, a PDF 2 per 50 pages, and research 30 to 300 depending on how much it reads; the pricing page lists every operation. Failed calls and cache hits are free, asking a question about a page adds nothing, and the playground needs no key.
Where does the 38/40 number come from?
40 prospecting searches run against the live playground on 20 Oct 2026 (and 13 Oct before it, with the same result), with no paid key, from one location. Results move with the web, so we date every number and re-measure it.