MiroFish's 72,000 Is a GitHub Star Count, and Running It Requires a Paid LLM API Key

The artifact is a 38.83-second vertical clip, 1080x1920 at 9:16, in an MP4 container (ffprobe reports mov,mp4,m4a,3gp,3g2,mj2), video in HEVC Main as hvc1 at 473,105 bps and 30 fps across 1,165 frames, audio in HE-AAC at 44,100 Hz stereo and 96,344 bps, 576,391 bps overall, 2,797,899 bytes on disk. TikTok's metadata puts the post at 2026-10-07 17:18:43 UTC from uploader ai.amirtech, channel nickname "Amir | AI Tools & Tech", with the audio credited as "original sound" by "Amir | AI Tools, Tech" and a 330-character description. At capture on 2026-10-08 the counters read 5,831 views, 259 likes, 17 comments, 93 reposts, and 300 saves. I read all 19 extracted frames and all 153 transcript words. Every frame is a horizontal split. The lower half holds one unchanging setup: a bearded presenter in a white camp-collar polo behind a light blue MacBook, lit against a blue gradient backdrop. The upper half rotates through B-roll. Frame 1 shows the MiroFish README scrolled to "Our Vision", "Live Demo", and "Screenshots". Frame 2 shows the project's demo site: "Predict Anything" in large type over "The World's First Simple and Universal Open Source Swarm Intelligence Engine", buttons reading "Try Live Demo" and "★ Star on GitHub", and a handwritten line over a doodled crowd reading "Join the waitlist to get notified at launch!". Frame 3 is a stock shot of a woman at a ring light. Frame 4 is a dense Chinese-language node graph. Frame 5 is an AI-generated man in a suit in front of a PRC flag. Frame 7 shows four flashcards labeled "the product", "launch post", "new price", and "the ad". Frames 10 and 11 are a beige panel of scattered dots. Frame 12 is a Gource-style repository history animation stamped "Sunday, 21 December, 2025 02:23:15", with nodes reading MiroFish_logo.jpeg, MiroFish_logo_compressed.jpeg, PROJECT_STATUS.md, env.example, README.md, run.py, /backend, /frontend, /scripts, /services, action_logger.py, and a contributor avatar labeled "BaiFu". Frame 13 is a dark card reading "GOVERNMENT / THINK TANK" above "Policy Opinion Forecasting" and "Simulate public reactions before policy announcements to predict opinion trends and risk points across different groups". Frame 14 is the blue betta-fish logo beside the wordmark "MiroFish". Frame 15 shows three numbered cards: "1 free", "2 open source", "3 predicts the reaction". Frame 16 is three dot clusters around a document icon. Frame 17 is the README "Overview" with an "English | 中文文档" toggle. Frame 18 is a delayed stock quote table (107.38 / -0.63 / -0.58% / "11 Minutes Delay"). Frame 19 is a "comment Fish" call to action over the presenter.

The 72,000 figure counts GitHub stars, and the repo now shows 77,118

The narration says "72,000 people have already started using it." The GitHub REST API for 666ghj/MiroFish returns 77,118 stargazers, 11,788 forks, 460 watchers, and 136 open issues as of 2026-10-08. A star is a bookmark on a repository page. It records that someone clicked a button, which is a different act from cloning the project, installing Python 3.11, supplying two API keys, and running a simulation.

The project publishes its own star history chart in the repo at static/image/star-history-light.svg, last regenerated by commit chore: update Star History (#832) on 2026-10-01. Its x-axis runs from 29 Nov 2025 to 01 Oct 2026 and its y-axis tops out past 60k, which puts the 72,000 the video cites inside the plausible stargazer range for late September 2026. The number itself is real. The noun attached to it is wrong.

For comparison, panews reported 18,000 stars within days of 7 March 2026, when MiroFish reached number one on GitHub trending, and "over 22,000 stars" at the time of writing. The curve since then is steep, which makes star counts an unreliable stand-in for adoption.

"Completely free" covers the license, not the compute

The video says the tool is "completely free." The repository is licensed AGPL-3.0, so the source is free to download, read, modify, and self-host. The cost shows up one layer down. The README's Quick Start requires an LLM_API_KEY for any OpenAI-SDK-compatible endpoint and recommends qwen-plus through Alibaba's Bailian console. Alibaba's Model Studio documentation for qwen-plus lists a 1,000,000-token context window with max input of 997,952 tokens and max output of 32,768, priced per token: "Input: 0.115 Per 1M tokens, Output: 0.287 Per 1M tokens" for inputs under 128k in the China (Beijing) region, rising to 0.689 per 1M input tokens between 256k and 1M.

A simulation that runs thousands of agents with long-term memory through multiple interaction rounds is a token-heavy workload by construction. The README also requires a Zep Cloud key for memory and states that Zep's "free monthly quota is sufficient for simple usage", which is the project's characterization rather than a measured figure for a thousand-agent run.

AGPL-3.0 also carries an obligation the video does not mention. If you modify MiroFish and expose it over a network, the license requires you to offer the modified source to your users.

The hosted product is a waitlist, not a live service

Frame 2 captures the line "Join the waitlist to get notified at launch!" on the project's own demo site. Pulling the site's JavaScript bundle at 666ghj.github.io/mirofish-demo/assets/index-BnqFA15a.js confirms a full waitlist implementation: WaitlistForm, waitlistAlreadyJoined, waitlistDuplicate, waitlistChangeEmail, and the string "waitlist to get notified at launch!".

So the paths available to a founder watching this video are self-hosting the AGPL code with your own keys, or joining a list. Search results surface several other MiroFish-branded domains advertising hosted monthly plans. The repository's homepage field points to mirofish.ai, and the only MiroFish web destination linked inside the README is the github.io demo. Anything else charging a subscription is not linked from the project.

The project's own site lists marketing as one of six use cases, and never mentions pricing tests

The video frames MiroFish as a launch-testing tool and closes on "new product launches, pricing tests, and ad campaigns." The demo site's bundle contains the complete use-case list, six entries, each with a tag and an example:

• Financial Decision Support (Finance / Investment), example "Market sentiment simulation after central bank rate hikes"
• Policy Opinion Forecasting (Government / Think Tank), example "Carbon tax public acceptance simulation"
• Crisis PR Simulation (Enterprise / PR), example "Product recall sentiment simulation"
• Marketing Strategy Testing (Brand / Marketing), example "New product launch social media reaction prediction"
• Story and Fiction Deduction (Entertainment / Creative), example "Predicting the lost ending of Dream of the Red Chamber"
• Academic Research Support (Academic / Research)

One of those six matches the video's pitch. The rest aim at governments, investors, PR teams, novelists, and social scientists. The README's two featured demo videos are a Wuhan University public opinion forecast and a prediction of the lost ending of Dream of the Red Chamber. "Pricing tests" appears nowhere in the README or the site bundle. The claim that founders are already running pricing tests through it has no first-party source I could find.

The 10-day build is real, and the project has shipped for ten months since

The video's "20-year-old student in Beijing built this in just 10 days" holds up. panews names Guo Hangjiang, 20, a senior at Beijing University of Posts and Telecommunications, who built MiroFish in 10 days. 36Kr reports that Chen Tianqiao of Shanda Group committed 30 million RMB, roughly 4.1 million USD, within 24 hours of seeing a demo video. The GitHub profile for 666ghj lists the name "BaiFu", company "USTC, BUPT", and location "Shanghai, China", where he moved after the funding. The README's recruiting address is mirofish@shanda.com, which corroborates the Shanda connection.

The 10 days describe the first push. The repo was created 2025-11-26T04:23:02Z, took its AGPL license on 2025-12-19, and has shipped v0.1.0 (2025-12-22), v0.1.1 (2026-01-22), and v0.1.2 (2026-03-07), with commits through 2026-10-01. The Gource animation in frame 12, stamped 21 December 2025, is a snapshot from week four of a project that has now been under active development for ten months.

The simulation core is also not original work. The README credits OASIS (Open Agent Social Interaction Simulations) from CAMEL-AI as the engine. That repository was created 2024-11-13, is Apache-2.0, and carries 5,240 stars. MiroFish adds GraphRAG seed extraction, a web frontend, and report generation on top of it.

Key Takeaways

• Verified: MiroFish exists at github.com/666ghj/MiroFish, licensed AGPL-3.0, written in Python, created 2025-11-26, with 77,118 stars and 11,788 forks per the GitHub API on 2026-10-08.
• Verified: Guo Hangjiang, 20, a BUPT senior, built the first version in 10 days, and Chen Tianqiao's Shanda Group invested 30 million RMB, per panews and 36Kr.
• Correction: The 72,000 the video cites tracks GitHub stars, which are bookmarks, rather than people who have "started using it."
• Correction: No first-party source supports "pricing tests" as a MiroFish use case. The project's own site lists six scenarios, one of which is marketing strategy testing with a product-launch example.
• Partial correction: "Completely free" applies to the AGPL-3.0 source. Running it requires a per-token LLM key, with Alibaba listing qwen-plus at 0.115 USD per 1M input tokens and 0.287 USD per 1M output tokens at short context.
• Clarification: The hosted version was a waitlist at capture time, confirmed by the waitlist form code in the demo site's JS bundle.
• Clarification: Guo studied at BUPT in Beijing, and his GitHub profile now lists Shanghai, where Shanda is based.
• Context: The simulation engine is OASIS from CAMEL-AI, an Apache-2.0 project created 2024-11-13. MiroFish layers GraphRAG, a frontend, and report generation on top.
• Context: The video omits that a 30 million RMB funded company sits behind the "free" repository, with a recruiting address at mirofish@shanda.com in the README.

Resources

• 666ghj/MiroFish on GitHub establishes the AGPL-3.0 license, the 2025-11-26 creation date, the 77,118 stars and 11,788 forks as of 2026-10-08, the Qwen and Zep key requirements, the OASIS credit, and the mirofish@shanda.com recruiting address.
• The MiroFish live demo site establishes the "Predict Anything" hero, the waitlist state, and the six documented use cases with their category tags and examples.
• panews on Guo Hangjiang establishes the founder's full name, his age of 20, his BUPT enrollment, the 10-day build, and the 18,000-star count around the 7 March 2026 GitHub trending milestone.
• 36Kr's report establishes Chen Tianqiao and Shanda Group as the investors, the 30 million yuan figure, and the 24-hour decision.
• Alibaba Cloud Model Studio: qwen-plus establishes that the README's recommended model is live, metered per token, and tiered by context length.
• camel-ai/oasis establishes the prior art: an Apache-2.0 multi-agent social simulation framework created 2024-11-13, a year before MiroFish.