NeuroArxiv skill
Before designing a new mechanism the agent checks real arXiv papers, reads them in isolation and converges on the findings.
- 438 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The skill makes network requests to arXiv and triggers many agent calls.
- The installer writes files into agent skill directories.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/neuroarxivDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill makes network requests to arXiv and triggers many agent calls.
- The installer writes files into agent skill directories.
This is third-party code. Review the repository files before installing.
What it does
NeuroArxiv stops an agent before it starts writing from scratch. The agent maps the task onto arXiv categories and search terms, fetches papers over real HTTP, then runs a separate isolated read per paper, typically 10 to 20. Results are scored, clustered and converged into a recommendation with known weaknesses of each approach. In Claude Code the skill is called with `/neuroarxiv "<problem>"`, in Codex by name. Without an explicit call the agent judges whether a search is worthwhile: for renaming a variable or a CRUD form it declines and goes straight to code.
Who it is for. Engineers designing caching, ranking, protocols or ML components who want to rely on published work.
Good fit when
- A consensus or coordination scheme must be chosen and a ready approach may have been published.
- A naive ranking version will break at scale and known limits matter.
- The agent is about to build a key architecture component that is costly to redo.
- A cited work should be checked for withdrawal or narrow applicability.
Not a fit when
- The task is routine: a form, an SDK wrapper or a one-off script.
- The user already picked the algorithm or library.
- A broad web search is needed rather than arXiv.
Example request
/neuroarxiv "cache invalidation scheme for a distributed service" — find arXiv papers, read each in isolation and name the known weak spots.Limitations
The skill is token-expensive: a real arXiv fetch and about one isolated agent call per paper. Search covers arXiv only, so its reach is narrower than general web search, which the README admits in its comparison. The result stays a recommendation, so open and verify the cited papers. The installer needs Node 18 or newer. The last commit dates from September 2026. Installation downloads the package from GitHub through `npx`.
How to disable. Delete `~/.claude/skills/neuroarxiv` and `~/.codex/skills/neuroarxiv`, or disable the plugin if it was installed as one.
FAQ
Which agents are supported?
Claude Code and Codex CLI. Without a flag the installer picks whichever is found on the machine, and `--all` installs into both.
Why does the agent sometimes decline?
It first checks three questions: is there a technical mechanism, will real effort be invested and did the user leave the approach open. An explicit `/neuroarxiv` call skips the check.
How does it differ from ordinary search?
Each found paper is read in isolation and the answer names a weakness of that specific paper. In the README comparison with web search such flags appeared in 5 of 5 problems.
Where is the evaluation data?
In `EVALS.md` and the transcripts in `bench/deep-tech-eval-transcripts.md`; the authors also report cases where the method lost on citation breadth.
Is the skill expensive to run?
In tokens, yes: a real arXiv fetch and about one isolated agent call per paper. That is why the agent declines routine tasks unless you call the skill explicitly.
Discussion
No questions or reviews yet.
Related
SkillsAn autonomous iteration skill for agents: modify, verify, revert failed changes, and repeat the loop toward a target metric
A skill for second-agent review of plans or code: Claude Code starts rounds, Codex reviews, and they exchange files
autoresearch drives an agent through an experiment loop with a fixed benchmark and budget, committing only measured improvements.
A library of 30-plus instruction skills for coding agents: planning, building, testing, hallucination guards and shipping
An architecture coach skill: questions lead to ADRs, a diagram and a plan, and a repository is audited from its code
Two skills for coding agents: wtf audits a project before shipping, better-codex keeps the agent disciplined while working on code
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| NeuroArxiv (this card) | no | 438 | |
| arXiv Review Paper Harness | no | 487 | leads to a survey article for arXiv in LaTeX; NeuroArxiv is a prior-art check before designing |
| arxiv-to-prompt | no | 172 | turns a paper into a prompt; NeuroArxiv finds and reads a set of papers for the task |
README in short
The skill makes an agent read arXiv first: categories, isolated reading of each paper and convergence. The authors put the five-problem evaluation in `EVALS.md`.