Research Agent Skills
Turns any coding agent into a research co-scientist that searches literature, verifies citations and drafts papers.
- Source repo
- KalarisLabs/research-agent-skills
- Stars
- ★ 76
- Last updated
- today
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 61/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeChatGPT · OpenAI API · Claude API (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A PhD student drafting a first paper: scientific-writing for the IMRaD draft, unslop-academic-writing to strip AI phrasing, citation-verification to check every reference.
- Not a fit if
- Teams whose work must stay fully offline without Crossref/OpenAlex lookups
- Researchers who want a GUI web app rather than a terminal coding agent
- Non-academic writers who need no citation checking or paper formats
- Source review
- 61/100 · Some gaps
What does this agent do, and when should you use it?
Research Agent Skills is an open-source library from Kalaris Labs of Agent Skills that follow the agentskills.io specification: each skill is a plain folder holding a SKILL.md playbook, reference notes and tested scripts. The README claims 281 skills spanning hypothesis generation, literature search and PRISMA systematic reviews, experimental design and statistical power, data analysis with field-standard libraries (scanpy, rdkit, astropy, pydicom, pytorch-lightning and others), scientific writing and figures, citation verification, journal formatting, arXiv submission and peer review. Skills load on demand — the agent sees one short description each and opens the full playbook only when a task matches, so an installed library costs little context. One npx skills add command installs it into Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and roughly 70 harnesses, or into Claude Code as plugin bundles per field. The library ships no model; inference and cost come from whatever agent you already run.
Each skill folder contains a SKILL.md (trigger conditions, workflow steps, verification rules) plus runnable Python or CLI scripts. For citation-verification, the agent reads refs.bib or a manuscript's reference list and queries Crossref, OpenAlex and arXiv per entry, then reports each reference as verified, mismatched or not found. systematic-review-prisma covers protocol, search strings, deduplication, screening and the PRISMA 2020 flow diagram and checklist; literature-review searches PubMed, arXiv, bioRxiv and Semantic Scholar and extracts study data. Analysis skills such as statistical-analysis, pydeseq2, scanpy, rdkit and astropy wrap domain libraries in workflows with assumption checks and reporting conventions. unslop-academic-writing ships a linter plus a rewrite playbook for stock phrasing like "delves into" and "pivotal role". venue-templates and journal skills (nature-portfolio, cell-press, ieee-transactions) instruct the agent to read current author guidelines before formatting, and arxiv-submission checks TeX sources, .bbl files and figure paths. Outputs are Markdown or LaTeX manuscripts, BibTeX, figures and structured check reports.
- A PhD student drafting a first paper: scientific-writing for the IMRaD draft, unslop-academic-writing to strip AI phrasing, citation-verification to check every reference.
- A lab running a systematic review: systematic-review-prisma produces search strings, dedup, screening and a PRISMA 2020 flow diagram and checklist.
- An ML researcher submitting to NeurIPS, ICML or ICLR: ml-paper-writing supplies official LaTeX templates and arxiv-submission catches packaging errors before upload.
- A bioinformatician analysing single-cell data with scanpy or differential expression with pydeseq2, then writing the methods section from the same session.
- A reviewer or corresponding author: peer-review produces a bounded assessment draft and rebuttal-and-response-to-reviewers writes point-by-point replies under venue length limits.
- A librarian or research software engineer moving libraries with reference-manager-interop across Zotero, Mendeley, EndNote and LaTeX or Word.
How do you install or deploy this agent?
Requires Node.js 18+ and no account. The command reads the GitHub repository directly; the README states there is no npm package with this name.
Install every skill for every detected agent:
npx skills add KalarisLabs/research-agent-skills --allInstall specific skills or target one agent:
npx skills add KalarisLabs/research-agent-skills --skill citation-verification apa7
npx skills add KalarisLabs/research-agent-skills --agent claude-codeInstall a field bundle via the Claude Code plugin marketplace:
/plugin marketplace add KalarisLabs/research-agent-skills
/plugin install biology-research@research-agent-skillsAlternatives include GitHub CLI and a SHA-256-verified standalone installer:
gh skill install KalarisLabs/research-agent-skills
curl -fsSL https://raw.githubusercontent.com/KalarisLabs/research-agent-skills/main/install.sh | shHarnesses without skill support can read skills/<name>/SKILL.md directly, or you can copy the folder into the harness's skills directory (often .agents/skills/).
How do you use this agent?
After installing, ask the agent in plain language; it selects the matching skill from its description. For example:
Verify every reference in refs.bib and tell me which ones don't exist.Run a PRISMA 2020 systematic review on exercise and depression: search strings, dedup, flow diagram.Analyze this scRNA-seq dataset with scanpy and write the methods section.Update installed skills later:
npx skills updateNote that Crossref, OpenAlex, arXiv and PubMed lookups need network access, and some skills (Firecrawl Research Index, Exa, OpenRouter image generation) require your own provider keys.
What are this agent's strengths and limitations?
- Unusually broad coverage: 281 skills across literature, statistics, single-cell, cheminformatics, clinical imaging, physics and venue formats, so one install serves the whole research pipeline.
- Built on the open Agent Skills specification — plain folders with SKILL.md, no SDK or server — and movable across roughly 70 harnesses without vendor lock-in.
- Targets concrete failure modes with committed evidence: citation-verification reports precision and recall of 1.00 on 18 labeled references, and unslop cuts the slop index 83% in a pilot, with chart scripts and result files in the repo.
- On-demand loading keeps the resident context cost near zero: the agent sees one short description per skill until it is needed.
- CI gates on every pull request: validate, skill-security and benchmarks workflows plus OpenSSF Scorecard and Best Practices badges.
- Model-based results are pilots only — 23 tasks judged by one small model — and are not independent validation.
- Skill depth varies substantially; the repo reports a mean static quality score of 95.5, but long-tail library skills have no third-party reliability check.
- Many skills depend on external services and keys: Crossref, OpenAlex, arXiv, PubMed, Firecrawl, Exa and OpenRouter image generation all need network and often credentials.
- Repository metadata lists the license as NOASSERTION while the README badge says MIT, so verify the LICENSE file before commercial use.
- It is a collection of prompts and scripts, not an orchestrated pipeline: there is no end-to-end runner or guarantee, and the user remains responsible for correctness.
How does this agent compare with similar options?
The README contrasts these skills with a plain AI assistant: an assistant drafts what you ask for, while the skills push the agent through a research workflow that checks its own output against citation databases, reporting guidelines and current venue rules. It also differs from vendor-specific prompt packs by following the open agentskills.io specification and installing through the skills.sh CLI.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Research Agent Skills This agent | 61 · Some gaps | CLIFree | ★ 76 | today | Python | Codex · Claude Code |
| OpenDraft | 81 · Good | CLIFree + model costs | ★ 521 | 13d ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| PaperOrchestra Skill Pack | 62 · Some gaps | Agent plugin / skillFree + model costs | ★ 694 | 18d ago | Python | Claude Code |
| PaperJury | 79 · Good | Agent plugin / skillFree + model costs | ★ 1.2k | 2mo ago | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md documents least-privilege workflows (permissions:{} with per-job grants), script egress allowlist, SHA-256 verification plus Sigstore/SLSA attestations, tar extraction rejecting absolute paths and '..', and a private vulnerability reporting process — all statically evidenced. Deductions: install paths include curl|sh and irm|iex pipe-to-shell, a high-risk default; --yes skips confirmation; skill scripts run with agent permissions and the docs only advise users to review and sandbox, with no enforced confirmation gate; sensitive-data handling is limited to 'put API keys in env vars' with no data classification or redaction rules; rollback relies on 'npx skills update' and doctor tamper detection with no explicit uninstall/revert procedure; external effects (Crossref/OpenAlex calls) are described generally but not enumerated per skill.
README and installation docs are largely consistent on commands, bundle names and harness lists, and the CI workflows (benchmarks, code-security) match the controls the docs describe, so self-consistency is decent. Dependency availability only states Python 3.9+/3.11+ and uv with no lockfile or offline guidance; failure messages are limited to SECURITY response SLAs, with no error-handling or messaging conventions for skill scripts, hence low scores.
Targets students, PIs, systematic reviewers, ML/AI, bio/chem/medicine/physics researchers across a very broad scenario set, and supports ~70 harnesses including Claude Code, Codex, Cursor and Gemini CLI, so environment fit is strong. Capability boundaries are stated (local files, some scripts need API keys, tool permissions vary by harness) but framed marketing-style without per-skill limits; trigger precision depends on agent auto-routing and the hit@3 0.96 figure is self-reported and unverifiable statically, so 2.
Information architecture is clear (README TOC, catalog, docs sections, FAQ, quickstart), install notes are thorough with multiple methods and a parameter table, and examples/FAQ are rich. However the LICENSE file is MIT while metadata says NOASSERTION, an inconsistency; versioning is limited to pyproject 1.0.0 and RAS_VERSION with no CHANGELOG; known limitations appear only inside the fly-brain study, not as a project-level limitations doc; maintenance responsibility points to Kalaris Labs but the publisher is unverified, so 2.
Output usability has concrete task examples and structured outputs (verified/mismatched/not found) for citation-verification and unslop, but actual output quality for most skills cannot be verified statically. Marginal value comes from 280+ skills spanning the whole workflow, though the volume risks context bloat (the FAQ itself admits installing everything loads every description); cost-benefit is zero-cost install with on-demand loading, offset by hidden costs of scientific package dependencies and API keys, so 2 across.
README references benchmarks/results, KEY_CLAIMS.json and CI chart-drift checks, giving reasonable claim traceability. But headline numbers (95.5/100, hit@3 0.96, 21/23 task wins) are self-reported and depend on paid model runs, so cross-source corroboration is only 1. Fact/inference separation is decent: the fly-brain scope statement distinguishes simulation predictions from living-tissue claims, and verified/UNVERIFIED values are separated.
- Install docs offer curl|sh and irm|iex pipe-to-shell methods, a high-risk default; prefer npx skills or verify signatures before executing.
- Skill scripts run with agent permissions and the docs only advise user review and sandboxing, with no enforced confirmation or permission minimization.
- LICENSE is MIT but metadata says NOASSERTION; confirm actual licensing before use.
- Benchmark figures (quality scores, routing hit rate, task win rate) are self-reported and depend on paid model runs, not independently verifiable statically.
- Publisher is unverified, so maintenance responsibility and update-path reliability cannot be confirmed from the files.
- No CHANGELOG or project-level known-limitations doc; version rollback and uninstall procedures are unclear.
FAQ
Does it cost anything?
Do I have to use Claude Code?
Can citation-verification really catch fabricated references?
Does it work on Windows?
irm ... install.ps1 | iex, described as verifying pinned release downloads with SHA-256.