AI Research Skills (Orchestra)
Orchestra-Research/AI-Research-SKILLs98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle.
With caveats
Broad access Depth 4/5
Scientific Agent Skills Claude skills
K-Dense's library of 177 science skills covering bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing.
Our verdict Recommended with caveats
Deep, specialist and actively maintained, but install only the skills you need and check each skill's license and external-service requirements.
What Scientific Agent Skills does
Scientific Agent Skills is a large library from K-Dense Inc. that gives a coding agent working procedures for scientific computing. The 177 skills span biology and genomics (Biopython, bulk RNA-seq, deepTools, cellxgene census), chemistry and drug discovery (DeepChem, DiffDock, datamol), medicine and clinical work (clinical reports, clinical decision support), physics and Earth science (Astropy, Cantera), general data tooling (Dask, AnnData) and communication (citation management, scientific slides, peer review). A database-lookup skill folds in dozens of public databases, and several skills wrap commercial or institutional systems such as Benchling, DNAnexus and Adyptv.
The skills are thick: SKILL.md files average about 14 KB, many with reference folders, and a good share ship Python helper scripts (sample-sheet validators, setup checks, assay pipelines). Each skill declares its own license and version metadata, and the README warns that these licenses may differ from the repo's MIT license. The repo describes blog benchmarks for a few skills, but our snapshot has no automated test suite. An autoskill skill watches repeated research workflows through a local Screenpipe screen-history daemon and drafts new skills. Install by copying or symlinking the repo into an agent's skills folder; the README lists commands for Cursor, Claude Code and Codex, among others.
Good for
- Researchers in biology, chemistry or clinical data who want an agent to follow field-specific procedures
- Picking a few relevant skills (for example Biopython, RNA-seq, citation management) out of a large catalog
- Teams already using K-Dense's tooling
Not for
- Installing all 177 skills at once into one agent (descriptions alone will crowd the context)
- Clinical decisions, since the pack's own safety notes say output is not clinical advice
- Anyone who needs one uniform license across all skills
What's inside depth 4/5
177 skills averaging about 14 KB with reference folders and many helper scripts (2,780 files), but no automated test suite in the repo snapshot.
- Skills
- 177
- Reference docs
- 1129
- Scripts
- 536
- Tests / evals
- 19
- Templates
- 191
- Avg SKILL.md
- 13.6 KB
Counted from the repository file tree. How we score depth.
Dependencies and costs
- Python environments with the scientific packages each skill names (uv is suggested)
- Per-skill accounts or API keys for services such as Benchling, DNAnexus or Adyptv
- Screenpipe daemon and a local LM Studio or cloud LLM for the autoskill skill
Security review
Broad accessThis skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
- Pre-approves broad shell commands23 skills set
allowed-tools: Read Write Edit Bash(unrestricted shell, some with Agent, WebFetch or WebSearch); examples are aeon, alphagenome, arbor, autoskill, citation-management and database-lookup.skills/aeon/SKILL.md:5
About 40 scientific-computing skills, mostly instructions plus Python helpers and API wrappers, where 23 pre-approve unrestricted Bash. The autoskill skill reads your passive screen-capture history from Screenpipe and sends redacted summaries to a local model by default, or to Anthropic/Foundry if you opt in. Other skills call public science APIs or vendor APIs with your own keys; most of the repository's files, including binary fixtures, were not available to us.
Worth knowing
- Makes network requestsautoskill reads your Screenpipe screen/audio history from a local API on user request and, with the default local backend, sends redacted app and window-title cluster summaries to a local LM Studio server; the opt-in claude and foundry backends send those summaries to api.anthropic.com or a Foundry gateway.skills/autoskill/scripts/run.py:449
- Needs an API key for its own serviceANTHROPIC_API_KEY, FOUNDRY_API_KEY, LM_API_TOKEN and SCREENPIPE_TOKEN are each sent only to their own service; remote plaintext HTTP endpoints are refused.skills/autoskill/scripts/backends.py:103
- Makes network requestsOn request downloads the BIDS schema and BEP catalogue from readthedocs.io and raw.githubusercontent.com (HTTPS only, size-capped, validated) and overwrites reference files in the skill folder; data, not instructions.skills/bids/scripts/update_schema.py:36
- Makes network requestsUses the Exa SDK with the user's EXA_API_KEY and adds an x-exa-integration attribution header; other skills call PubMed, OpenAlex, Crossref, DepMap, EBI, AlphaGenome, Adaptyv, Benchling and similar public or vendor APIs.skills/exa-search/scripts/exa_search.py:95
- Needs an API key for its own serviceEXA_API_KEY, NCBI_API_KEY, OPENALEX_API_KEY, ADAPTYV_API_KEY and similar keys are read from the environment and sent only to the matching service.skills/exa-search/scripts/exa_search.py:95
- Installs with curl | shREADME suggests installing uv with
curl -LsSf https://astral.sh/uv/install.sh | shfrom the uv vendor's own host.README.md:321 - Installs packages or toolsSkills instruct
uv pip installof scientific packages (mostly with version pins) into separate environments; autoskill also downloads a sentence-transformers model on first use and suggests building screenpipe from source.skills/autoskill/SKILL.md:100 - Relies on a CLI or package we could not reviewArbor can use the optional arbor-agent CLI and git worktrees; citation-management may call parallel-cli; other skills use cellprofiler, DiffDock, Screenpipe, Benchling, DNAnexus tools and similar software that is not in the repository.skills/arbor/SKILL.md:6
What we checked
The repository (2780 files) on Oct 29, 2026. Not a full review: 2173 of 2780 files were not downloaded (large repository; we reviewed the skill folders); it contains compiled binaries or archives we cannot read; we read 11 of 81 code files; it relies on code outside the repository: screenpipe daemon and LM Studio used by autoskill; arbor-agent CLI and parallel-cli referenced by arbor and citation-management; PyPI packages installed by the skills (alphagenome, exa-py, scholarly, ete4, sentence-transformers and others).
NVIDIA SkillSpector static scan: 316 pattern hits (Data Exfiltration 188, Agent Snooping 49, Dangerous Code Execution 25, Excessive Agency 14, Tool Misuse 9, Privilege Escalation 9). The two CRITICAL taint hits are autoskill/doctor.py sending the user's own Screenpipe or LM Studio token to the configured local endpoint for a health check. The curl-pipe hit is the uv installer line in the README; most Data Exfiltration, Tool Misuse and Privilege Escalation hits are API-key, kubeconfig, DELETE-endpoint and --nocheck mentions in reference documents, and the 'obfuscation' hits are a GFP protein sequence in a DiffDock example.
A review, not a certification. Skills change after we check them — read the files before installing.
License
Repo is MIT, but each skill declares its own license in its SKILL.md and the README says these can differ; check the ones you install.
Can it run unattended?
Not really. Individual skills describe analysis procedures and scripts, but the pack has no shared loop, stop conditions or resumable state for unattended runs.
Alternatives to Scientific Agent Skills
ToolUniverse
mims-harvard/ToolUniverseHarvard MIMS lab's Python SDK, MCP server and 556 skills that give agents access to thousands of scientific databases and tools for biomedical research.
Recommended
Broad access Depth 4/5
ScienceClaw
beita6969/ScienceClawA research agent system built on the OpenClaw gateway whose Skills and typed Operators are updated only after a replay-verified execution; ships with 319 skills.
With caveats
Check first Depth 3/5
Hugging Face Skills
huggingface/skillsOfficial Hugging Face skills led by hf-cli, a Hub CLI reference generated from your local install, plus AWS/SageMaker deployment and operations skills.
Recommended
Check first Depth 3/5
FAQ
What is K-Dense Scientific Agent Skills?
It is an open-source library of 177 agent skills for scientific work, from K-Dense Inc. It covers bioinformatics, chemistry, clinical research, data analysis, lab integrations and scientific writing, and works with Claude Code, Codex, Cursor and other agents.
How do I install Scientific Agent Skills?
Clone the repo and copy or symlink the skills you want into your agent's skills directory; the README gives commands for Cursor (a local plugin symlink) and for Claude Code and Codex. Because there are 177 skills, start with a small subset.
Are the Scientific Agent Skills free?
The repo is MIT licensed and free to use. Individual skills may require third-party accounts, paid APIs or local software, and each skill's own license field can differ from the repo license.
Is it safe to install?
We found no hidden network calls or credential theft, but the skills run code, install packages and call external services, and the README itself says to review what you install. The autoskill skill reads Screenpipe screen history, so enable it only if you accept that.
Sources: GitHub repository (stars, license, file tree, last push Oct 26, 2026). Review based on the README and SKILL.md files as published.