# An open-source project that cost less than a relay station, 32 days to 1,121 stars: the format is real, but half the heat was bought · GitHub open source

> Source: https://stringers.io/en/github/gh_hanzohanzhe__Grandet.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

> Relay prices are chaotic and everyone was looking for a neutral price-comparison site. It converted full listed prices and put them into agents; a large share of GitHub stars were bought, and the inflated growth rate then drew a round of aggregator reposts.

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

hanzohanzhe/Grandet · hanzohanzhe

# An open-source project that cost less than a relay station, 32 days to 1,121 stars: the format is real, but half the heat was bought

**What it is** A price-comparison library for API relay stations, packaged as a skill and MCP that an agent can install with one sentence. It finds the cheapest endpoint and writes it into your config.

**Why it took off** Relay station pricing is messy, and people have been looking for a neutral comparison site. It converts all the list prices and loads them into agents. But a large share of its GitHub stars were bought, and the growth spikes drew a ring of aggregator-account reposts.

**C** **How you could follow this** Stop building yet another Chinese relay-station price comparison. Copy the format: take data nobody else wants to organize, turn it into an anonymous MCP plus a one-sentence-install Skill, and publish it to several skill marketplaces and MCP directories at once.

C grade: the "data site repackaged as an agent Skill/MCP" distribution approach is worth learning from. But seven or eight repos already compete on the same topic, half its stars were manufactured, organic demand is unproven, and the data needs paid weekly upkeep. Copying it wholesale isn't worth it. Circulating in small circles

Data (updated daily) · Repo created 2026-10-05 · 1,175 stars · 7 days +161

GitHub · stars

Data 2026-10-17 → 2026-11-08 (23 days, 23 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

Grandet scrapes list prices from roughly 1,894 domestic API relay stations, converts them by rate multiplier into price per million tokens, and packages the result as a one-sentence-install Skill for Claude Code and Codex, plus a no-signup MCP. It picked up 1,121 GitHub stars in 32 days, but almost all the accounts that forked it are bot-like, and its star curve abruptly broke on 10-31. Below: what's real, what's bought, and if you want a project that earns stars the same way, which part to copy and which to avoid.

### A question that had been asked for half a year

2026 year May 3: a user named dcsite posted on linux.do: "There are countless relay sites out there. So is there an aggregator site that specializes in reviews, price comparisons, and user comments on relay sites? Neutrality comes first. Some pad their numbers, some run off with the money, and some are unstable and go down every few days."

Nearly all the replies were cold water: "Anyone with the energy and motive to do this is either the relay station itself or someone with a stake." "Everyone's tried it, and none are neutral. These review sites take money and do the job." "Thankless work. Every site buys a plan to test, and by the time you publish, it's gone." Others spelled out how messy the business is: some "set up complicated exchange-rate rules so you can't work out the price," others "advertise a 0.3 multiplier and then play word games, or mark the real 0.8 in fine print."

5 and a half months later, in Oct 19, a new account called hanzohang replied with the 25nd comment under the thread: "I built one myself. Writing a crawler to pull prices from relay stations across the web isn't hard. The hard part is that every update of tens of thousands of api entries burns a lot of tokens to upload... It updates once a week for now... the site is grandet.ai." The reply got 1 upvotes.

The same day, its GitHub repo hanzohanzhe/Grandet gained 93 stars in a single day, its biggest day to date.

### What it actually is

A price-comparison library with a layer on top that lets agents use it themselves.

- **Data** : 1,894 sites, in three columns: official cash prices, inference-provider cash prices, and relay-station credit prices. Relay prices are converted to dollars per million tokens using "model multiplier × group multiplier × completion multiplier," with the original group name and observation date noted. USD credits, CNY credits, and cash are listed separately and not compared head to head (grandet.ai, 11-06).

- **Jargon glossary** : multiplier, group, official relay, reverse engineering, account pool, car-pooling, public-good station, check-in. 144 terms with Chinese-English equivalents.

- **Free tiers** : 75 official free-tier offers and 101 free relay-station groups, with notes on whether a card or real-name verification is needed.

- **Install into an agent** : tell your agent "Install Grandet Skill for me: [grandet.ai/download.md](https://grandet.ai/download.md)" and it installs itself. Supports Codex, Claude Code, Cursor, Gemini CLI, Cline, OpenClaw, and 24 other apps. The remote MCP is anonymous and needs no key. The bundled script can write a chosen relay station straight into the Claude Code / Codex config. It backs up first, and the change can be undone.

- **Neutral** : ranks only by list price. It never calls a site "stable" or "full-strength," and it never repeats a site's own marketing copy. That directly answers the linux.do complaint that "none are neutral."

The author's GitHub account was registered 2022 years ago, has 28 followers, and this is its only repo. Their linux.do account was registered just 10-18 ago, with the signature "强取胜于苦耕" ("seize, don't toil").

### How the stars grew

Start with the numbers (GitHub, day by day):

| Date | star | fork |
|---|---|---|
| 10-17 | 202 | 4 |
| 10-19 | 357 | 32 |
| 10-25 | 710 | 65 |
| 10-30 | 994 | 100 |
| 10-31 | 1,012 | 109 |
| 11-01 | 1,014 | 112 |
| 11-06 | 1,121 | 121 |

From 10-17 to 10-30 it steadily gained 40 to 90 stars a day, about 56 on average. Then on 10-31 it dropped to 18,11-01, just 2. Real things rarely grow like this. It looks much more like a two-week plan expiring.

The fork accounts tell more. GitHub hides this repo's star list, so you can't see who starred it, but forks are visible: among 128 fork accounts, over 80% of the 115 with public repos have nothing but forks; 66 forked the same batch of unrelated small repos. I spot-checked several: all had 0 followers, jumbled names, clustered registration dates, and each had starred exactly 100 repos. Fork timing matches the star timing: it starts on 10-16 and ends on 10-30.

So of the 1,121 stars, the roughly one thousand added after 10-16 are, by inference, largely bought. The exact number can't be verified. About 10-16 stars came before 117, and I couldn't trace where they came from.

### What the bought stars bought

The manufactured growth was picked up by a ring of automated accounts that watch GitHub growth: on Instagram, SteadyVane wrote that it went "from 28 to 117 in 249 hours" and was "6.3 times faster than other repos it tracks"; AISignal (10-19), hellobuilder, findthe.tools, a CSDN gitblog repost (10-27), a YouTube Short "Grandet: Never overpay for LLM APIs," and LinkedIn reposts.

This is all machine-driven amplification, with weak real-person spread. A blogger with 100k followers on X posted an introduction that got 1,257 views and 3 likes 10-20; I found no discussion on V2EX, Reddit or Hacker News.

### Where real users came from

If real people are using it, they most likely arrive through agents, not from the GitHub homepage.

- The same Skill is published to GitHub Releases, ClawHub (OpenClaw's skill marketplace), and SkillHub, and it's listed in the skill directories of mcpservers.org, Glama, and TypingMind.

- The homepage's first line is written for AI assistants ("For AI assistants: ..."), and there's an llms.txt. Each model has its own price-comparison page, such as "claude-sonnet-5 API comparison: cheapest relay stations," and long-tail Chinese keywords like "cheapest Claude API" and "relay station multiplier" surface it in search.

- The latest Skill package has been downloaded 4,264 times (GitHub, 11-06). The site itself reports 13,241 Skill downloads and 7 price lookups over the past 18,159 days. The scale is plausible, but these are the author's own stats and unverified.

In other words, roughly half of the GitHub stars were staged for aggregator accounts. The real distribution is "agent searches → one-sentence install → used to check prices."

### Who it's competing with

Relay-station price comparison and navigation was a crowded topic on Chinese GitHub in the second half of 2026:

| Project | Launched | star(11-06) | How it differs from Grandet |
|---|---|---|---|
| zzsting88/relayAPI | 2026-03 | 5,443 | Hand-written recommendations and reviews, not full price data |
| dimthink/PriceAI | 2026-05 | 5,112 | Compares subscription resale prices, not API prices |
| qixing-jk/all-api-hub | 2025-09 | 5,048 | Browser extension that manages your own relay-station balance and check-ins |
| Folkbench/folkbench | 2026-10-02 | 4,312 | Ranks by measured availability and latency; its growth pattern is also suspicious (single-day +561, very few forks) |
| panxunying/ai-coding-welfare | 2026-08 | 1,277 | Navigation for relay stations and free tiers, with one-click config |
| Fei2411/real-api-pricing | 2026-09 | 1,071 | Calculates token unit prices for subscription plans |
| liuda30201/AI-Coding | 2026-10-15 | 608 | Navigation + setup scripts |

Others include tokencost (official prices only), OpenRouter / pricepertoken (no relay sites), and a few that test whether relays are legit: Veridrop, api-relay-audit. Since October, new ones keep appearing: nervecenter (10-23), awesome-coding-plan (11-03). One more detail: the same batch of accounts that boosted Grandet also forked other repos on the same topic, so several of these may have used the same star-farming channel.

What really sets Grandet apart is just one thing: **full-price listings + multiplier conversion, packaged as a Skill/MCP that an agent can install directly and use to edit config** . Others either benchmark quality, maintain manual lists, or only track official prices.

### Its weaknesses

- **The data has to keep flowing.** The author admits each full refresh of tens of thousands of entries "burns a pile of tokens," and updates weekly; the main repo hasn't been pushed since 10-27. A price comparison library is worthless once it stops updating.

- **Neutrality can't be proven.** Long-time users on linux.do have already said it: someone doing this "has logically lost credibility." Grandet's terms say "if there's a partnership, it will be stated separately," which leaves room for referral commissions. Once money comes in, neutrality gets hard to claim.

- **It only looks at list prices, not quality.** The biggest headaches with relay sites are watered-down models, vanishing operators, and sites going up and down every few days. Grandet doesn't cover any of that; projects like Folkbench are starting to fill that gap.

### If it were you

**C: The distribution playbook is worth learning, but building another one on this same topic isn't worth it.**

**Who it suits** : people who already have a dataset nobody else bothers to organize (prices, free quotas, config parameters, a list of tools in some ecosystem) and want more developers to use it and maybe pick up stars along the way. Best in a different niche: LLM inference price comparisons in English or Japanese, GPU rental prices, or something in an agent ecosystem nobody has summarized yet.

**Who it doesn't suit** : anyone wanting to build another Chinese relay-site comparison/navigation (at least seven or eight repos on the same topic are already competing, several with equally dubious stars); anyone who can't keep the data updated weekly; anyone counting on star farming to drive growth (fork accounts are easy to expose, and a cliff like 10-31's is visible to everyone).

**How to do it** : copy the format, not the topic. Put the data on a website and keep only the Skill and MCP in the repo. Install by telling the user to "paste this line to your agent." The MCP is anonymous and needs no key. It can help users write results straight into config, backing up first and allowing rollback. Put a line for AI assistants at the top of the homepage and provide llms.txt. Bilingual README. Filling the gap Grandet left (quality testing, uptime) will differentiate you far more than another price list.

**How to get users** : publish the same Skill simultaneously to GitHub Releases, ClawHub, SkillHub, MCP directories (mcpservers.org, Glama), and TypingMind; make a searchable page for each entry to catch long-tail queries; reply on forum threads where people are actually asking, but don't expect the first reply to bring much traffic (Grandet's top reply got 1 likes). Stars have to come from real growth that triggers the "trending" aggregator accounts on their own.

**Timing** : Skill and MCP directories are still early, so inclusion bars are low and ranking near the top isn't hard. Once skill marketplaces start sorting by install count and the top spots get locked in, this tactic gets much more expensive. Rough estimate: a few months left (inferred).

**What would show you misread the situation** : (1) if Grandet keeps gaining 20+ stars a day organically after star farming stops and keeps updating weekly, the natural demand for this topic is bigger than it looks, so raise the estimate; (2) if a skill marketplace publicly shows install counts and Grandet ranks high, it means "agents as a distribution channel" already brings more users than the GitHub homepage does, and you should move fast; (3) the reverse: if Grandet stops updating before January and the site data goes stale, maintenance cost has outweighed the payoff, and data projects like this need a clear plan for who keeps feeding them data long-term.

[GitHub repo ↗](https://github.com/hanzohanzhe/Grandet)

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

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| OpenRouter / pricepertoken.com / benchlm.ai | Web | 2023-2025 | OpenRouter, a large platform; pricepertoken covers 300+ models | Only official prices and legitimate inference providers. Doesn't include Chinese relay sites, doesn't convert multipliers or groups, and can't be installed into an agent to edit config. |
| AgentOps-AI/tokencost | Open source | 2023-12 | 2,003 stars | A library for calculating official prices, with no relay data and no agent integration. |
| Folkbench/folkbench | Open source | 2026-10-02 | 4,312 stars (single-day peak +561; our database flags "very few forks, a new account's single-repo surge") | Benchmarks relay uptime and latency and ranks them; Grandet only looks at list prices, covers more, and is packaged as a Skill/MCP. |
| zzsting88/relayAPI | Open source | 2026-03-09 | 5,443 stars | Hand-written relay recommendations and reviews, not a full price database. |
| panxunying/ai-coding-welfare, liuda30201/AI-Coding | Open source | 2026-08-21 / 2026-10-15 | 1,277 / 608 stars | A navigation for relay sites, charity sites, and free quotas; no full price comparison and no MCP. |
| dimthink/PriceAI, bytedoger/awesome-OpenPrice | Open source | 2026-05 / 2026-07 | 5,112 / 1,081 stars | Compares card-reseller prices for subscriptions, not API token prices. |
| Fei2411/real-api-pricing | Open source | 2026-09-05 | 1,071 stars | Calculates per-token equivalents of subscription plans; doesn't include relay sites. |
| mn-api/awesome-ai-proxy | Open source | 2025-02 | 1,002 stars | A static list of relay sites. |
| Relay arenas: aiapipk.com / Veridrop / api-relay-audit | Web | 2026 | Small (Veridrop has 725 open-source stars; api-relay-audit has 906) | Tests whether relays are legitimate and their uptime; has sponsored placements; doesn't do full price comparisons. |
| qixing-jk/all-api-hub | Open source | 2025-09 | 5,048 stars | A browser extension that manages balances and check-ins on relay accounts you've already registered; no price comparison. |

Where users come from

- **Verified** **Buying stars/forks (botnet of accounts)** : of 128 fork accounts, 115 have ≥80% of their forks be forks themselves; 66 of them also forked the same batch of unrelated small repos (hestjs, BullFlow, ai-native-callcenter, AgnesAI-Models, etc.); spot-checked accounts have 0 followers, sign-up dates cluster together, and each starred exactly 100 repos; forks and stars grew at a steady rate from 10-16 to 10-30, and both fell off a cliff at 10-31. The stargazer list is hidden by GitHub, so how many stars were bought can't be directly verified.

- **Verified** **Re-posting by GitHub growth aggregator accounts** : SteadyVane (Instagram: 28 hours 117→249 stars), AISignal 10-19, hellobuilder, findthe.tools, aiindigo, CSDN gitblog 10-27, a YouTube Short, and a LinkedIn undercode-news post are all automated or semi-automated reposts.

- **Verified** **X creator** : Dan Kornas (100K followers) 10-20 intro post, only 1,257 views and 3 likes (fxtwitter)

- **Verified** **Skill marketplace / MCP directory distribution** : 10-16 was published simultaneously to GitHub, ClawHub, and SkillHub; mcpservers.org, Glama, and TypingMind skills all list it; installation is just telling the agent one sentence.

- **Verified** **SEO and SEO for agents** : a comparison page per model, FAQs, Q&A pages, llms.txt, and price-finder.md; a line for AI assistants at the top of the homepage; Chinese long-tail keywords are indexed.

- **Verified** **linux.do community** : the author's account hanzohang (10-19) replied once under an old "neutral price comparison site" thread, which got 1 likes, and never posted a topic; a LINUX DO friend link was added to the README.

## Grandet (hanzohanzhe/Grandet): why it took off

> In one sentence: it scrapes prices from 1,894 sites in China's gray-zone "API relay" market, converts them to per-million-token prices using multipliers and groups, and packages the result as an Agent Skill plus an anonymous MCP, so Claude Code / Codex can find the cheapest endpoint in one sentence and write it into your config. 32 days, 1,121 stars on GitHub. But **a large share of those stars and forks were bought** : fork accounts are almost entirely bot accounts, 10-16 to 10-30 steadily gained about 55 stars a day, then 10-31 suddenly dropped to single digits. What holds up is the topic (relay price comparison) and the "install into your agent" distribution format, not natural GitHub popularity.

Data as of: 2026-11-06. Sources are cited in each section.

### 1. Basic data

- Created 2026-10-05, author hanzohanzhe (GitHub account registered 2022-12, 28 followers, this is their only repo); the same-named linux.do account hanzohang lists the location as "cambridge" with the signature "Strength comes from seizing, not from toiling," registered 10-18 (linux.do profile page).

- Star curve (our github repo_day, with daily data from 10-17): 10-17 202 → 10-20 438 → 10-25 710 → 10-30 994 → 10-31 1,012 → 11-01 1,014 → 11-06 1,121. Single-day peak +93 (10-19). From 10-17 to 10-30 it averaged about +56 a day; from 10-31 it dropped to 2–26 a day.

- About 117 stars 10-16 ago: the Instagram aggregator SteadyVane posted "from 117 to 249 within 28 hours" (https://www.instagram.com/p/DemRtDiEUQE/).

- Forks 121 (11-06 live 125); fork-to-star ratio about 11%, looks normal; but fork accounts are almost entirely bot accounts (see section 4).

- Releases: skill-v0.3.0 shipped 19 builds between 10-14 and 10-18; the latest build.20261018b attachments total 4,264 downloads, build.20261016c 1,125, and 20261015e 925 (GitHub releases API).

- Self-reported by the site (grandet.ai header, Google snippet, exact date unknown): "Last 7 days: 7,899,258 requests · 3,456,188 visits (Cloudflare analytics); Skill downloads 13,241 · price lookups in the last 7 days 18,159." Much of the Cloudflare request and visit count may be crawlers or agent calls, so treat it only as weak evidence of usage; not verified.

- The main repo hasn't been pushed since 2026-10-27 (pushed_at 10-27). Since 11-06, a fork account, seojunee, has submitted over 50 neatly formatted issues and commits (looks like batch output from an agent), and the site's community section is full of "AI agent experience test" comments, suggesting the author used agents to run tests and feed the results back.

### 2. What the product is and where it differs

- **Data** : covers 1,894 sites in three columns: official cash prices, inference-provider cash prices, and relay-site credit prices. Relay prices are converted to per-million-token prices using New API's "model multiplier × group multiplier × completion multiplier," with the original group name and observation date listed; USD credits, CNY credits, and cash are shown separately and not mixed (https://grandet.ai/).

- **Glossary** : 144 terms including multiplier, group, 1:1, official relay, reverse-engineered, account pool, car-pooling, charity sites, and check-in, in English and Chinese (repo references/glossary.md).

- **Free quota library** : 75 official free quotas + 101 free relay-site groups, noting whether a card or real-name verification is required (https://grandet.ai/zh-CN/faq/).

- **Agent form** : one line, "Install Grandet Skill for me: [grandet.ai/download.md](https://grandet.ai/download.md)", and the agent installs it itself. Works across 24 apps including Codex, Claude Code, Cursor, OpenCode, Gemini CLI, Cline, Roo, Kilo, Windsurf, Hermes, OpenClaw, Trae, WorkBuddy. Remote MCP at `https://grandet.ai/mcp`: anonymous, no key. `connect.mjs` writes the chosen relay straight into the configs of Codex, Claude Code and others, backs up first, and can "undo". The same Skill is also shipped to ClawHub and SkillHub (commit 10-16 "Ship the same Skill to GitHub, SkillHub and ClawHub"; ClawHub page [clawhub.ai/hanzohanzhe/skills/grandet-api-price-finder](https://clawhub.ai/hanzohanzhe/skills/grandet-api-price-finder)).

- **Neutral stance** : ranks only by listed price, never calls any provider "stable" or "full-strength", and doesn't repeat the site's own hype copy. The terms say partnerships "will be disclosed separately if they exist", leaving room for affiliate or featured-listing revenue later. The site is also working on Grandet Personal, a desktop app (private beta), and enterprise consulting (data card recall reasons).

### 3. Competitors

| Competitors | Where | Launched | Scale | How it differs from Grandet |
|---|---|---|---|---|
| OpenRouter / pricepertoken.com / benchlm.ai / llmapiprices.xyz | Web | 2023~2025 | OpenRouter is the big platform; pricepertoken covers 300+ models | Only official prices and licensed inference providers; no domestic relays, no multiplier conversion |
| AgentOps-AI/tokencost | Open source | 2023-12 | 2,003 stars | A pricing calculation library for official prices; no relays, can't be installed into an agent |
| zzsting88/relayAPI | Open source | 2026-03 | 5,443 stars | Hand-written relay recommendations and reviews, not a full price database; no MCP |
| Folkbench/folkbench | Open source + folkbench.com | 2026-10-02 | 4,312 stars, peak +561 in a day (10-25); our database flags "very few forks, single-repo spike" | Measures uptime and latency of real relays for rankings and switching; Grandet only looks at listed price |
| dimthink/PriceAI, bytedoger/awesome-OpenPrice | Open source + website | 2026-05 / 07 | 5,112 / 1,081 stars | Compares card-reseller prices for ChatGPT/Claude subscriptions, not API token prices |
| panxunying/ai-coding-welfare, liuda30201/AI-Coding | Open source | 2026-08-21 / 10-15 | 1,277 / 608 stars | A directory of relays, public relays and free quotas, with one-click config scripts; no full price comparison, no MCP |
| Fei2411/real-api-pricing | Open source | 2026-09-05 | 1,071 stars | Computes subscription monthly fee ÷ actual usable tokens, comparing Coding Plans |
| mn-api/awesome-ai-proxy | Open source | 2025-02 | 1,002 stars | A static relay list |
| Relay arena aiapipk.com, Veridrop (veridrop.org, 725 open-source stars), api-relay-audit (906 stars) | Web / open source | 2026 | Small | Tests whether relays are legitimate and their uptime; has sponsored placements; doesn't do full price comparisons. |
| linux.do "One Tracker" | Community public-relay tool on the forum | 2025 | Small niche | Users enter multipliers themselves and compare by hand |
| qixing-jk/all-api-hub | Open source (browser extension) | 2025-09 | 5,048 stars | Manages balances and check-ins for relays you've already signed up with; no price comparison |

Data from our GitHub database (repo_profile, 11-06) and each site's homepage. Conclusion: "relay price comparison / directories" was a crowded Chinese-language GitHub topic in the second half of 2026, with several repos taking off in the same period. Grandet's edge is **full listed prices + multiplier conversion + packaged as an Agent Skill/MCP that writes configs directly** . Others either only test quality, are hand-curated directories, or only cover official prices.

### 4. Growth levers: where users (and stars) come from

#### 4.1 Bought stars / forks (strong evidence)

- GitHub hides this repo's stargazer list (REST returns 404, GraphQL totalCount=0), so star accounts can't be checked, only forks.

- I pulled all 128 fork accounts: **115** are accounts with "≥5 public repos, of which ≥80% are forks"; **66** forked the same batch of unrelated small repos, such as hestjs, BullFlow (jaikydota/BullFlow, 353 stars, created 10-14), ai-native-callcenter (rasonyang, 547 stars, 103 forks), AgnesAI-Models, ai-shortVideo-pipeline, GetClaudePro, Lyapunov, forkd. Spot-checked accounts (leathabobica, denamanodflapuen, alebobinloviegra, tstornez, chicf28) all have 0 followers, garbled names, and registration dates clustered together (2025-08-09, 2025-09-20, 2026-01-19, 2026-01-28, several each), each starred exactly 100 repos, and their public repos are all forks. A typical bot pool used for inflating metrics.

- Fork dates: 10-16 on 9, 10-17 on 15, 10-18 on 10, then 3–10 per day, continuing after 10-30,10-31 at 1–4 per day. The star pace is identical: 10-17–10-30 per day steady at +40–90,10-31, dropping to +18,11-01, then +2. **The start and end line up with a star-purchase package.**

- Rough estimate: of the roughly 1,000 stars added after 10-16, a substantial share were probably bought. The exact ratio can't be verified (stargazers are hidden).

#### Second-wave exposure from 4.2 viral posts (strong evidence, small scale)

- The inflated "star velocity" was picked up by aggregator accounts that auto-track GitHub growth: SteadyVane (Instagram, "6.3x faster than other repos it tracks"), AISignal (10-19, [www.aisignal.dev/analysis/hanzohanzhe-grandet](https://www.aisignal.dev/analysis/hanzohanzhe-grandet)), hellobuilder.ai, findthe.tools, aiindigo, mcpservers.org, Glama, TypingMind skills, toolpassed, a CSDN gitblog repost (10-27, [blog.csdn.net/gitblog_00813/article/details/161705930](https://blog.csdn.net/gitblog_00813/article/details/161705930)), a YouTube Short "Grandet: Never overpay for LLM APIs", and LinkedIn undercode-news.

- Dan Kornas (100k followers) posted an intro on X 10-20, but it only got 1,257 views and 3 likes (fxtwitter API). Searching "grandet" in our x-radar corpus returns 0 posts.

#### 4.3 Community (weak)

- The author 10-19 replied to an old linux.do thread, "Is there a neutral aggregator site dedicated to relay reviews, price comparison and commentary?", introducing grandet.ai. It got only 1 likes (https://linux.do/t/topic/2105674/25); this account has never started a topic. The same day, a LINUX DO friend link was added to the README (commit #20).

- No hits on V2EX, Reddit or HN.

#### 4.4 Agent/Skill ecosystem distribution (structural, likely the main real source of usage)

- The Skill ships to GitHub Release, ClawHub (OpenClaw's skill marketplace) and SkillHub. Installation is "paste one sentence to your agent"; the homepage's first line reads "For AI assistants: ……", and it provides llms.txt and price-finder.md, which are themselves SEO aimed at agent search and agent marketplaces.

- The site has a page per model (e.g. "Anthropic · claude-sonnet-5 API price comparison: cheapest relays"), an FAQ, and Q&A pages (/answers/…). Chinese long-tail terms like "cheapest Claude API" and "relay multiplier" are indexed on Google.

- The self-reported Skill downloads of 13k and weekly price lookups of 18k match the order of magnitude of release asset downloads (4,264), which suggests people are actually installing and using it. But the install counts come from the author's own tally and are unverified.

#### 4.5 what I didn't find

- No evidence of paid ads (this kind of repo wouldn't buy them); no media coverage; not on HN or Product Hunt; the repo is under 60 days old, too young for GitHub trending (trending only picks repos about 60 days old).

### 5. Why it charted

1. **Riding a trend** : in the second half of 2026, many people in China used Claude Code/Codex through relays, with wildly inconsistent prices (three sets of numbers for multiplier, grouping and recharge exchange rate). V2EX and linux.do were full of posts on "how to calculate the real price" and "pitfalls to avoid". "Is there a neutral price-comparison site?" was a question asked again and again on linux.do. Folkbench, AI-Coding, PriceAI and other repos on the same topic took off at the same time.

2. **A new form** : others built web pages or lists; this one is packaged as an Agent Skill plus anonymous MCP that installs with one sentence, and it can directly edit Claude Code/Codex configs. Timing matched the rise of the Skill and ClawHub ecosystem.

3. **Deep data** : 1,894 sites, multiplier conversion, sorted by currency, with observation dates and 144 pieces of jargon explained. A step up from hand-made directories and awesome lists.

4. **Much of GitHub's star count was manufactured** : the fork pool and the cliff at 10-31 both show that 1,121 stars shouldn't be read as organic interest. The inflation's real effect was pushing the repo into "star velocity" aggregators, which brought a round of secondary exposure.

### 6. Was differentiation why it took off?

The agent form plus full-coverage relay price conversion is the clearest differentiator in its niche, and it's the source of real downloads (release 4,264, self-reported Skill downloads 13k). **The GitHub star numbers, however, mostly come from inflation and have little to do with differentiation.** So the verdict: product differentiation is real, popularity is partly manufactured.

### 7. Visibility and copycats

- Visibility: circulating in small circles, mostly via machine aggregator accounts; no media coverage, no meaningful spread on X, revenue not disclosed (and there is no clear paid tier yet).

- Copycats: didn't see anyone directly cloning Grandet; many parallel repos on the same topic (Folkbench, liuda30201/AI-Coding, waw666waw666/awesome-ai-relay-and-cards all started in October, NerverDevs/nervecenter 10-23, supercli-site/awesome-coding-plan 11-03). Notably, the same batch of fork accounts that inflated Grandet also forked ThinkWatch, so several repos on this topic may have used the same inflation channel.

### 8. Takeaways

- Packaged a "data site" as an **Agent Skill + anonymous remote MCP + llms.txt** , installable with one sentence, turning agents into a distribution channel and using install receipts to track installs.

- Picked an opaque grey market with messy terminology, and used a neutral stance of "ranking by listed price only, no judging quality" to build full-coverage data, with monetization planned later through partnerships and affiliate fees (the terms already leave room for this).

- One Skill shipped to several skill marketplaces (GitHub Release, ClawHub, SkillHub, TypingMind, etc.), plus a bilingual Chinese-English README.

- The inflation playbook: a bot pool bought about 50 stars and forks per day for roughly two weeks, making the curve look organic and triggering automatic coverage from "star velocity" aggregators. The cost is that the fork accounts are easy to spot.

### 9. Uncertain

- How large the purchased share of stars really is (stargazers are hidden, so it can only be inferred from forks).

- Whether the site's Cloudflare visit counts and Skill download counts, which are self-reported, include crawlers or the author's own test agents.

- Where the roughly 117 stars before 10-16 came from; I couldn't find out.

- The author's identity (Cambridge?) and how far the monetization plans (affiliate fees, desktop app, enterprise version) have progressed.

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