LLM-MM-Agent's install command installs a different repository's package

The only shell command installs a skill pack from another account

MM-Agent ships as a Python codebase: MMAgent/ at the top level, a config.yaml beside it, twenty-five lines in requirements.txt, and a demo/ directory carrying its own backend and frontend. The single shell command on the page does none of that. It reads:

npx skills add luckyfan-cs/mm-agent-expert --skill '*' -a codex -a claude-code -a cursor -g -y

That pulls a bundle published under a different account, luckyfan-cs, and installs it into three coding clients instead of into a Python environment. The instructions right below the command agree with that reading: Node.js and npm have to be present, a new session is started afterwards, the user types the expert name to invoke it, and the problem with its data are handed over as a prompt. Compiling the resulting PDF needs a LaTeX engine in whatever runtime executes the session. So the first thing a newcomer copies sets up a prompt pack, and the agent code the rest of the page describes is never fetched by it.

Twenty-five dependencies, four version floors, no declared Python

requirements.txt names 25 packages. Four carry a lower bound: torch>=2.6.0, torchvision>=0.21.0, transformers>=4.45.2 and pyarrow>=15.0.0. The remaining 21 resolve to whatever the package index serves on install day, and that set holds both vendor clients, openai and anthropic. No line declares which Python version the code expects. There is also no comment column and nothing marked optional, so a plain install takes the whole modeling stack whether or not a run touches it: pymc, hmmlearn, pmdarima and statsmodels for probabilistic work, scikit-image for images, sentence_transformers for embeddings. One line reads helm, a name also carried by the Kubernetes release command, so where that line comes from is easy to misread. Report assembly comes from pypandoc, pylatex and tiktoken, and bcrypt sits in the list for password hashing, a web application concern rather than a modeling one.

The demo and the root package run two different web stacks

The May 2026 news entry describes an open-source demo built from a Next.js frontend and a FastAPI backend, with bring-your-own-key configuration and E2B sandbox support, and sends readers to ./demo/README.md. The root requirements file describes a different application. It carries streamlit and none of fastapi, next or an E2B client, and the demo keeps its own demo/requirements.txt, so one checkout holds two dependency sets and two front end frameworks. The example files confirm the split: demo/.env.example, demo/backend/, demo/frontend/ and demo/scripts/. That .env.example is the file to read before a key is pasted anywhere, and the E2B sandbox is the part to understand first, because this pipeline writes and runs Python and that description places execution inside a hosted sandbox rather than on the machine holding the checkout.

Hosted access went from a star list to a WeChat code to a dead link

The access story in the news log changes four times. The December 2025 entry offers service accounts drawn from the star list, with limited server capacity given as the reason. The January 2026 entry switches to an invitation code handed out through a WeChat group. A second January entry strikes the hosted demo at cdn.mmagent.top out and says in parentheses that it went away with the server, redirecting readers to the open-source demo. Only the May 2026 entry describes something a reader can obtain directly. One endpoint still carries no retraction: the HuggingFace Space MathematicalModelingAgent/MathematicalModelingAgent, linked from the October 2025 contest item. Each step moved the trial further from hosted and closer to self hosted, so evaluating this project now means running it on your own machine and supplying your own key.

A contest placement with no scoring script attached

The October 2025 item reports that two undergraduate teams took a Finalist Award in MCM/ICM 2025, placed as the top 2.0% of 27,456 teams. Nothing at the root backs that number with a measurement. The listing runs .gitattributes, .gitignore, LICENSE, MMAgent/, MMBench/, README.md, README_zh.md, __init__.py, assets/, config.yaml, demo/, doc/, docs/, figs/ and requirements.txt, with no scoring script, prompt log, sample dataset or judging note among them. A directory named MMBench sits beside the agent code, and the news log never says what it measures or which MMBench it follows. Read the placement as an outcome of one contest run reported by its authors, not as a measured property of the system.

The four stage list ends on an unbalanced heading marker

Under Technical Details the workflow is laid out as four numbered stages. The first one carries two bullets, covering the problem background, objectives, data availability and constraints, and the decomposition of a problem into subtasks. The second stage heading is printed with its bold markers left unclosed, and no bullet ever appears under it. The two stages named higher on the page, computational solving and solution reporting, get no lines of their own either. The same page promises automated model selection, interactive data analysis with visualizations, iterative code improvement and automatic paper writing, and none of it is tied to a flag, a file or a setting in the checkout. A config.yaml sits at the root next to pyyaml in the dependency list, and not one key inside it is named on the page, so what a run reads at startup stays opaque.

Two documentation directories and a package init at the root

The checkout carries both doc/ and docs/, plus a second README in Chinese beside the English one. Between them the file listing shows no index page, and demo/README.md is the only nested README named there. An __init__.py also sits at the repository root rather than inside MMAgent/, which lays the whole checkout out as an importable package and puts the directory name itself on the import path. Two venues are claimed for one paper, an AI4MATH workshop at ICML 2025 in July and NeurIPS 2025 in September, and both entries link the same arXiv identifier, 2505.14148. The tree carries a LICENSE and a GPL-3.0 grant, and no release artifacts sit next to it.

Editorial conclusion

Read the paper before the news log if you are planning a deployment. LLM-MM-Agent is a research release with no versioned artifact, an install command that belongs to another account, and a demo that needs your own key on your own machine. Check demo/README.md and the 21 unpinned dependencies against your environment first, and expect the second workflow stage to stay undocumented.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. README
4. usail-hkust/LLM-MM-Agent on GitHub