# A "free alternative to Exa" comparison chart drove 650 stars in two days: RustXNG and Google, the open-source project it deleted on day three

> Source: https://stringers.io/en/github/gh_sahenjp__RustXNG.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> The author posted a comparison chart on X showing "free local search hits 79, paid Exa 77" (556k impressions). A partner AI blogger amplified it, and AlphaSignal plus a batch of AI aggregator sites reposted it, hitting the pain point of per-call pricing for agent search.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

sahenjp/RustXNG · sahenjp

# A "free alternative to Exa" comparison chart drove 650 stars in two days: RustXNG and Google

**What it is** A free search service you run yourself on your computer or server. It queries dozens of search sources for you, then hands the results and page text to an AI agent. No API key, no per-query fee.

**Why it blew up** The author posted a comparison chart on X showing "local free search hits 79, paid Exa 77" (556K views). A partner AI blogger quoted it and amplified it, and AlphaSignal plus a batch of AI aggregator sites reposted it. It hit a real pain point: agent search billed per query.

**C** **How you could follow it** Don't build yet another SearXNG wrapper. Learn its playbook instead: pick a category developers pay for per call that can also run locally (OCR, transcription, content extraction), build a free single-binary + MCP version, and publish a fairly run comparison chart. Launch it yourself and get friends with followings to quote it.

Grade C: the cold-start playbook is worth learning, but the spot is already taken by SearXNG, xerj (19K stars) and several others. Its accuracy rests on scraping upstream search engines, which carries terms-of-service risk, and the "free alternative to Exa" claim has already been used, so copying it isn't worth the effort. What's circulating in small circles

Data (updated daily) · Repo created 2026-10-18 · 903 stars · 7 days +856

GitHub · stars

Data 2026-10-25 → 2026-11-08 (15 days, 15 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

A Japanese student developer rewrote the open-source meta-search engine SearXNG in Rust, packaged it as "free local search for AI agents," and posted a chart on X claiming hit rates close to the paid Exa. It gained about 650 stars in two days. On day three, someone pointed out that its accuracy came mainly from scraping Google. The author removed Google and publicly apologized, and the hit rate dropped noticeably. Below: why it could break out, where the stars came from, how thin its foundation is, and how you'd get the same stars.

### A post saying "search APIs are about to lose their jobs"

On the evening of Nov 1, 2026 (Japan time), @sahenjpn posted on X in Japanese: 「やっべ、検索APIの存在意義を揺るがしてしまったかもしれない…」(Uh oh, I may have just shaken the reason search APIs exist…), with a comparison chart: on the same 100 questions, his RustXNG got 79 right with its first result, paid Exa got 77, and Tavily got 34.

The post got 556K views, 2,689 likes and 3,357 bookmarks on X (11-06). More bookmarks than likes means "saved for later, for when I build an agent."

An hour-plus later, Japanese AI blogger @studio_yebisu (about 17K followers, self-described "AI blogger who specializes in scaring people") quote-posted: "A student in his 10s built this and open-sourced it" and "There's no going back to the era before RustXNG," adding that 1,000 results came back in 3.3 seconds. The post also disclosed that he and the author are building another product, DSH (DeepSeek Harness), and he sent traffic to its beta. This one got 184K views and 3,033 bookmarks. Together the two posts reached about 740K views, but it was really acquaintances boosting each other.

The author's X account was created in 2026 September and had about 6,100 followers when the post went up. There are only two repos on their GitHub.

### What it actually is

A search service you run on your own computer or server. You send a keyword, it asks dozens of external search sources, merges and ranks the results, and returns JSON for your agent. It can also be plugged in directly as an MCP server for tools like Claude and Cursor, with just two tools: `search` and `read_page` (fetches the full text of a result page, with length capped by token count).

The pitch is one line: **No API key, no per-query fee, queries never pass through someone else's cloud.** A single Rust binary of about 17MB, available for Linux, macOS and Windows, plus a Docker image. Licensed under AGPL-3.0.

The README is clear about it: it has **no web index of its own** . Search results come from external sources. For local search you have to crawl a few sites yourself or import Wikipedia.

### Where the stars came from

| Date | star | Added that day |
|---|---|---|
| 10-18 | Repo created | — |
| 10-25 to 10-31 | 22 → 40 | Single digits per day |
| 11-01 | 47 | +7 |
| 11-02 | 258 | +211 |
| 11-03 | 695 | +437 |
| 11-04 | 776 | +81 |
| 11-05 | 815 | +39 |
| 11-06 | 844 | +29 |

(GitHub daily stars, 11-06; live on the day's page about 860 stars and 66 forks)

The sequence went like this:

1. **11-01, first on X's Japanese circle** : the author's own comparison chart, plus a partner's quote.

2. **11-02, English AI news takes over** : AlphaSignal published "RustXNG Rewrites Python's SearXNG in Rust, Cuts Query Time 10x," with alphasignal tracking parameters in the link. news.the9.pro wrote the same day that "the chance of finding the right page is as high as paid Exa search." Over the next few days, a wave of AI aggregator and auto-writing sites churned out articles in Japanese, Traditional Chinese, Simplified Chinese, Vietnamese and Italian, nearly all restating that 79 vs 77 chart.

3. **11-03, the controversy day was also its biggest gain** . Someone pointed out that it "is just scraping Google, ignoring the terms of service" ("普通にスクレイピングしてるだけ定期"). The author removed the Google-related code that day and posted an apology: "I got carried away and published it without checking the terms much" ("調子に乗って規約などをあまり確認せずに公開してしまいました"). That apology post also got 47K views. A wave of "go easy on the newcomer" voices appeared in the Japanese circle.

Nothing turned up on HN, and no main-subreddit post was found. No ads were bought. All the stars came from social posts and news reposts. After 11-04, growing by dozens a day is the natural fall-off after a burst.

The author kept busy: in two weeks they shipped up to v0.30, with over 118 PRs, and several outside contributors are proposing substantive changes.

### How thin its foundation is

That chart is all its promotional material, and also its biggest weakness.

- **Much of the accuracy is borrowed** . "Close to Exa" relied heavily on scraping upstream sources like Google and Yahoo! JAPAN. After removing Google and switching to only sources whose terms allow it, the default config's Hit@5 (whether the correct page is in the top 5 results) on the same document question set dropped from 84 to 67–69 (project README, 11-05 measured).

- **The numbers are self-tests** . The README now says itself that the question set was involved in development, some local indexes cover the relevant documentation sites, and the comparison with commercial APIs was not run at the same time, so "don't infer from this that it's better than Exa or Tavily."

- **It now splits into three configs** : Standard uses only 34 sources whose usage conditions have been recorded and that need no key; Clean only queries the local index and makes no outbound requests; Full adds sources with terms restrictions or unverified status. The README's words: "Full is not a switch for bypassing licenses," and the risk is yours to carry.

A Rust rewrite with an MCP wrapper is something others could build in days. It took off because of the "free, about as good as Exa" claim, not because of a technology no one else could match.

### Who else is doing the same thing

| Project | Form | Scale | How it differs from RustXNG |
|---|---|---|---|
| SearXNG | Open source, Python, from 2021 | 39k stars | The original, with the widest range of sources, but no MCP or content extraction built specifically for agents, and heavier to deploy |
| mcp-searxng | Open-source MCP, 2024-12 | 1.3k star | Just an MCP wrapper around SearXNG, and you have to run a separate SearXNG alongside it |
| xerj | Open source, Rust, mid-2026 | 19K stars, +1.5K in a week | Also a Rust single binary + MCP, but leans toward local indexing of code and docs, and is much larger in scale |
| Exa | Paid API | exa-mcp-server 5.2k star | Its own index, billed per query. This is the target RustXNG used as its benchmark |
| Tavily | Paid API | tavily-mcp 2.5k star | About $8 / 1000 queries, results already organized for the model |
| Brave Search API | Paid, with a free tier | MCP 1.6k star | Its own index, clear terms |
| Firecrawl / Jina Reader | Open source + paid service | About 199K / 12K stars | Focused on extracting page content, with search as a side feature |
| Built-in web browsing in Claude / ChatGPT / Codex | Big-company features | — | Works out of the box. Even @studio_yebisu admits that people who only use Codex or Claude may not feel the need for it |

(GitHub,11-06)

The "free self-hosted search for agents" spot already has several occupants, and xerj is twenty times its size. On GitHub it currently has only same-named forks, and no independent copycat projects have appeared.

### If it were you

**Verdict** : C. The cold-start playbook is genuinely worth learning, but chasing this project itself isn't a good use of time. The same spot is already crowded with SearXNG, xerj, mcp-searxng and others. Its accuracy rests on scraping upstream search engines, which runs into terms of service as soon as it scales. Its biggest surge of stars came from one self-test chart plus a round of acquaintances boosting each other, and when a second person uses the "free alternative to Exa" line, both the novelty and the credibility are diluted.

**Who it suits** : developers who already have clean data nobody else has, or a vertical index (say, technical docs in one language, regulations for one country, or a certain kind of paper), who can write Rust or Go, and who have some followers on X themselves or among friends.

**Who it doesn't suit** : anyone planning yet another "scrape a few more search engines and wrap them in MCP" project. That's the fifth SearXNG wrapper, and the terms problem lands on you just the same. Also anyone with no posting channel who expects GitHub to grow stars by itself (it had no chance of reaching trending in 19 days, so all its stars came from posts).

**How to do it** , ranked by how worthwhile:

1. **Copy the playbook, swap the target** : pick something developers pay for per call every day that can also run locally (OCR, speech transcription, web content extraction, vectorization, PDF parsing). Build a free, local, single-binary version with MCP built in, and benchmark it against the paid leader on the same questions. This time, hold out the test set from the start, run both side by side, publish the script, and leave no gaps for someone to pick apart.

2. **Fill the part it did poorly: clean local indexes** . Its Clean config currently can only query sites you crawled yourself and Wikipedia. Build a ready-to-use, regularly updated, properly licensed vertical index package (for example, "official docs for the major programming languages and frameworks") that agents can query without going online. You can release it on its own, or submit it as a PR to RustXNG to build a reputation.

3. **Switch language circles** : the Japanese X circle reacts strongly to "a student aged 10 built something that undercuts paid APIs." Chinese and Korean dev circles also eat this narrative, but only if you have a strong comparison number behind it.

**How to get users** : one chart, one line. The chart is your hit rate versus the paid leader on the same queries; the line is "no key, no cost, runs locally." Post the first tweet yourself, have a friend with followers quote-tweet within an hour, and stuff the README and hashtags with agent, MCP, RAG, grounding. News sources like AlphaSignal and a wave of auto-generated sites will come find you.

**When to move** : not urgent, but the sooner the better. The MCP wave in agent toolchains is still hot, but the "free alternative to Exa" angle has already been used once. Better to pick a different category target and ship before Q1 2027.

**Signs you're reading it wrong** : ① If RustXNG keeps stars above 50 a day for a month after dropping Google and Docker pulls pick up, people clearly want the "local + MCP" shell and don't care much about accuracy. Then the first path has a lower bar than described here, and following it is worth more. ② If Exa, Tavily or Brave launch large free tiers, or built-in web search in LLMs becomes the default, the "save on API costs" pain point goes away. ③ If someone makes the same comparison chart with clean data and gets thousands of stars, the category has room, and you should move faster.

### Data and sources

GitHub: sahenjp/RustXNG repo page, README (v0.30.0), PR #83, Issue #104, daily stars (10-25 to 11-06); stars on comparable repos (11-06). X: @sahenjpn 11-01 launch post and 11-03 apology post, @studio_yebisu 11-01 quote post (impressions and bookmarks as of 11-06). Reposts: alphasignal.ai, news.the9.pro 11-02 article, plus aggregators like hermes-ai.net, aisignal.dev, ssdnodes. Hatena Bookmark: 23 bookmarks. The "self-tested" numbers have no independent reproduction; actual Docker pull counts were not checked.

[GitHub repo ↗](https://github.com/sahenjp/RustXNG)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| SearXNG | Open source | 2021 | 39.3k stars | Python original, most complete engine, but no MCP or body extraction designed for agents, heavier to deploy |
| mcp-searxng | Open source | 2024-12 | 1.3k stars | Just an MCP wrapper for SearXNG; you have to run SearXNG separately |
| xerj | Open source | 2026 (about 129 days ago) | 19k stars, +7k in 1.5 days | Also a Rust single-binary AI search + MCP, leaning toward local indexing of code and docs, larger in scale |
| Exa | Other | MCP 2024-11 | exa-mcp-server 5.2k stars, commercial API | Self-built index, pay per call; RustXNG uses it as the comparison target |
| Tavily | Other | MCP 2025-01 | tavily-mcp 2.5k stars, commercial API | About $8/1000 per call, results already processed for the LLM; RustXNG is free to self-host |
| Brave Search API/MCP | Other | MCP 2025-06 | 1.6k stars | Has its own index and clear compliance, but paid / quota-limited |
| Firecrawl / Jina Reader | Open source | 2024 | 199k / 12k stars | Focuses on extracting body text; search is not the core |
| Claude/ChatGPT/Codex have built-in web search | Big-company features | — | — | Works out of the box, no setup needed; RustXNG only matters to people building their own agents |

Where users come from

- **查实** **X 作者首发（日语）** :@sahenjpn 2026-11-01 帖「検索APIの存在意義を揺るがしてしまったかも」附 Exa 对比图：55.6 万曝光、2,689 赞、3,357 收藏（X API）

- **Verified** **X KOL mutual promotion** : @studio_yebisu (17K followers) 11-01 a quote post with 184K views and 2,043 likes, stating they worked with sahen on DSH and are driving traffic for them (X API)

- **Verified** **AI newsletter** : AlphaSignal published 11-02, link tagged utm_source=alphasignal; the9.pro published the same day

- **Verified** **AI aggregators / SEO auto-writing sites** : hermes-ai, aisignal.dev, reviewramble, ssdnodes, fyve.co.jp, Youzhao, Vietnamese / Italian sites and more, 11/2–5 batch posts (Exa search)

- **Verified** **Controversy-driven second wave of exposure** : 11-03 was criticized for scraping in violation of terms of service; the author deleted the Google post and apologized (the deleted post got 47K views), and stars rose by +437 that same day

- **Verified** **Hatena Bookmark** : 23 bookmarks, mostly in 11/2

## Why RustXNG (sahenjp/RustXNG) took off

**In one line** : a Japanese student developer rewrote SearXNG in Rust and packaged it as "free self-hosted search plus body extraction for AI agents (JSON API / MCP)." In Nov 1 they posted a comparison chart on X claiming local search accuracy approaches the paid Exa. The post got 556k impressions. A team contact, Japanese AI blogger @studio_yebisu, then quote-posted it (184k impressions). AlphaSignal and a batch of AI aggregators and SEO sites followed with reposts, and GitHub gained about 650 stars in two days. The product hits the pain point of pay-per-call agent search. But its moat is thin: the core capability comes from scraping upstream search engines. On day three it was criticized for violating terms of service, the Google implementation was removed, and the hit rate dropped.

### 1. Data (source: our github library repo_day; GitHub API 2026-11-06)

- 2026-10-18 repo created. At 10-25 it had 22 stars, and by 10-31 only 40, staying flat for a long time.

- 11-01 47 stars → 11-02 258 stars → 11-03 695 stars (+437 in one day) → 11-04 776 → 11-05 815 → 11-06 844 stars (GitHub API live: 861 stars, 66 forks).

- Exposure was concentrated in 11/1–11/3, after which growth fell to only 30–80 stars a day. A typical pattern of a single burst followed by natural decay.

- Releases came fast: v0.30.x within two weeks (automated via release-please), with more than 118 PRs. External contributors TKY-27, waylake, maebahesioru and CopenDeCamp submitted substantive PRs.

### 2. Growth tactics: Japanese-language X launch → KOL quote → English AI news / aggregator reposts

| Timeline | Event | Evidence |
|---|---|---|
| 10-24/25 | A little early interest: someone bookmarked it on Hatena Bookmark (yuiseki 10/25) | b.hatena.ne.jp, 23 bookmarks in total, mostly in 11/2 |
| **11-01 13:08 UTC** | Author @sahenjpn posted "やっべ、検索APIの存在意義を揺るがしてしまったかもしれない…" with a precision comparison chart against Exa | X API: 556k impressions, 2,689 likes, 3,357 bookmarks, 274 reposts |
| **11-01 14:24 UTC** | @studio_yebisu (17k followers, "professional AI surprise maker") quote-posted: "10-year-old students build this and open-source it" and "can't go back to a world without RustXNG," 1000 results in 3.3 seconds. The post notes they are co-building DSH (DeepSeek Harness) with sahen and driving traffic to the DSH beta | X API: 184k impressions, 2,043 likes, 3,033 bookmarks |
| 11-02 | AlphaSignal published "RustXNG Rewrites Python's SearXNG in Rust, Cuts Query Time 10x," link tagged utm_source=alphasignal. The same day the9.pro published "finds the right page as often as Exa's paid search" (citing the project's self-test: Top1 hit rate 79/100, Exa 77, Tavily 34) | alphasignal.ai, news.the9.pro |
| 11-02~05 | A batch of automated / SEO content sites published in bulk: hermes-ai.net, aisignal.dev, reviewramble, ssdnodes (RustXNG vs SearXNG, in Japanese / Traditional Chinese / English), fyve.co.jp "Claude Code for Business," Youzhao (Chinese), Vietnamese and Italian sites, DeepWiki, reporank | Exa search results |
| 11-03 | Some in the Japanese AI scene plugged it into "Muse" for comparison tests (@29meat_ai, 19k impressions), and another person compared small-model retrieval against Brave (@rS_alonewolf) | X API |
| 11-03 | **Criticism emerged** : it was called out for simply scraping upstream sources like Google and ignoring terms of service ("普通にスクレイピングしてるだけ定期"). The author removed the Google-related code and publicly apologized ("調子に乗って規約などをあまり確認せずに公開してしまいました"). The deletion post got 47k impressions. The controversy itself brought another wave of exposure. | X API; PR #83 |
| 11-05/06 | v0.28 / v0.30 split into three configs: Standard uses only sources the terms allow, Clean queries only the local index, Full at your own risk. Standard's Hit@5 dropped from 84 to 67–69 | Issue #104, README |

Conclusion: all growth came from content and social spread. **No ads, no HN (Algolia finds nothing), no Reddit found** . The seed was the author's own "approaching Exa" comparison chart (6,103 followers, account registered only 2026-09 ago), plus a KOL quote from a team contact, which counts as a contact-driven boost. The "10-year-old student" persona and the "undercuts paid APIs" headline hit the sweet spot for the Japanese AI X circle. English AI news sites and AI auto-writing sites then spread it globally.

### 3. Competitors

| Name | Form | Launched | Scale | How RustXNG differs |
|---|---|---|---|---|
| SearXNG | Open source (Python) | 2021 | 39.3k stars | The original: the most features and the largest community, but no MCP or body extraction designed for agents, and heavier to deploy than a single binary |
| mcp-searxng (ihor-sokoliuk) | Open source MCP servers | 2024-12 | 1.3k stars | Just SearXNG exposed as MCP; you still have to run a separate SearXNG |
| xerj-org/xerj | Open source (Rust) | About 2026-06 | 19k stars, +7k in 1.5 days | Also a "Rust single-binary AI search + MCP," but leaning toward local indexing of code and docs, and larger in scale |
| Exa / Exa MCP | Paid API | 2024-11 (MCP) | exa-mcp-server 5.2k stars | Has its own index, pay per call; RustXNG's marketing uses it as the target |
| Tavily / tavily-mcp | Paid API | 2025-01 (MCP) | 2.5k stars | About $8 / 1000 calls (the9.pro), results already processed and fed directly to the LLM |
| Brave Search API / MCP | Paid API (with a free tier) | 2025-06 (MCP) | 1.6k stars | Has its own index, clear compliance |
| Firecrawl / Jina Reader | Open source + SaaS | 2024 | 199k / 12k stars | Focuses on extracting body text; search is secondary |
| Agent-Reach | Open source | 2026 | 109K stars | Focuses on social platforms and the whole web, broader positioning |
| LLMs ship with built-in web search (Claude / ChatGPT / Codex) | Big-company features | — | — | The built-in one needs no setup; even KOL posts admit "people using Codex / Claude may not feel they need this" |

### 4. Why it took off

1. **Real pain, sharp framing** : people building their own agents pay per search (Exa / Tavily), and free plans are slow or capped. "Runs locally, no API key, no per-call fees, and returns the page body too" is instantly clear.

2. **A screenshot-ready number** : self-tested on 100 questions, Top1 hit rate 79 vs Exa 77. One comparison chart was the whole promo kit, and almost every repost repeated it. (The numbers are self-reported, not run side by side, with no held-out test set; the author later softened the claim.)

3. **Persona and insider KOL push** : "10 coding for students" plus "genius" was amplified by partner @studio_yebisu. The two posts together drew about 740k impressions and about 6,400 bookmarks (more bookmarks than likes, meaning people saved it to use later).

4. **Rode the MCP / agent hype** : tags like mcp, rag, grounding, and ai-agents were everywhere, which AlphaSignal-style newsletters and many AI auto-writing sites scraped as news.

5. **The controversy didn't stop it, it caused a second wave** : after being accused of violating terms of service, the author immediately removed Google, apologized, and split it into three configs. Japanese-language circles spoke up in defense of the newcomer, and the repo gained the most stars that day (11-03).

### 5. Differentiation

Versus SearXNG: a single Rust binary (about 17MB, low memory), built-in MCP (search / read_page tools, output length controlled by tokens), optional fetching of body / PDF / JS pages, plus SSRF protection and robots.txt compliance. Versus Exa / Tavily: free, self-hosted, queries never go through a third-party cloud. **But the differentiation is not a moat** . "Accuracy close to Exa" comes largely from scraping upstream sources like Google and Yahoo! JAPAN, which is exactly what drew the criticism. Once Google was removed, Standard's hit rate dropped noticeably (Hit@5 84 → 67–69). It took off on the "free and about as good as Exa" **claim** and the spread, not on technology nobody else has. A Rust rewrite and an MCP wrapper can be copied in days (xerj is already a bigger version of the same thing).

### 6. Visibility and copycats

- Visibility: **talked about in small circles** . About 740k impressions on X's Japanese AI circles. English AI newsletters like AlphaSignal and many AI aggregator / SEO sites also wrote about it. But I found nothing on HN or the main Reddit subs, no mainstream tech press. It's an open-source project with no revenue.

- Copycats: GitHub has a batch of same-name forks (TKY-27, zapabob, sitne, mikuteto-dev, etc.). They're forks, not independent knockoffs. No independent clone spotted so far. Rust self-hosted agent search already existed (xerj), so this isn't a copycat wave.

### 7. Takeaways

- For developer tools, targeting a paid market leader and publishing a "free can match it" comparison chart is the cheapest cold-start content.

- On Japanese-language X: the author's own launch post plus a fan-base KOL quote (which also funneled traffic to his other product, DSH), pushing a new repo to a few hundred stars in two days.

- Stuffing the README and tags with agent / MCP / RAG / grounding gets it picked up in bulk by AI news and auto-writing sites, carrying the reach overseas.

- Cautionary lesson: "accuracy" built by scraping upstream carries terms-of-service risk. Once it blows up and people ask questions, you have to dismantle your own feature. Get the compliant mode right before launch.

### 8. Uncertain

- Actual usage is unknown: stars aren't deployments, and Docker pull counts weren't checked.

- Whether reputation and growth hold after removing Google and unvetted sources needs another week or two of watching.

- How many stars each channel contributed on November 2–3 can't be separated (the stargazers endpoint returned 404 this time, so no hourly data).

- "Exa 77 / Tavily 34" are the project's self-reported numbers, not independently reproduced.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.