sciClaw

sciClaw — your paired scientist

An AI research agent that lives on your computer, not in someone else’s cloud.

sciClaw runs locally. Your data stays on your machine. You bring your own AI—OpenAI, Anthropic, Groq—and switch anytime. Message it from Telegram or Discord, and everything it produces lands in a single ~/sciclaw folder you can open, version, and audit.

brew tap drpedapati/tap && brew install sciclaw
Getting started Read the docs

Free and open source. MIT licensed. Runs locally. Your data never leaves your machine.

Pedapati, E. (2025). sciClaw: A Lightweight Paired-Scientist Agent for Reproducible Biomedical Research Workflows. Preprint—coming soon.

Our philosophy Security Creator

Say as little or as much as you want

Start with a quick question. Get more specific when you need to. sciClaw meets you wherever you are.

1 A quick question — just like asking a colleague
“What do we know about TDP-43 in ALS?”

sciClaw searches PubMed, pulls recent papers, and summarizes the key findings.

2 Ask for a specific output — a document, a format, a file
“Draft a methods section and give me a Word doc with tracked changes.”

Produces a .docx you open in Word, with insertions and comments you can accept or reject.

3 Point to your data — attach files, reference datasets
“Summarize the demographics in participants.csv and flag any missing data.”

Reads the file, produces summary statistics, and highlights gaps before you start your analysis.

4 Specify the exact method — statistics, language, parameters
“Run Spearman correlations between P300 amplitude at Pz and ALSFRS-R decline in R. Correct for multiple comparisons with Benjamini-Hochberg at FDR 0.05. Use erp_data.csv.”

Writes and runs the R script, reports effect sizes, and saves the output with full provenance.

5 Hand it a full protocol — multi-step, reproducible, audited
“Follow analysis-protocol.md. Pull the latest data from REDCap, run the preprocessing pipeline in Python, then the stats in R. Output a Quarto manuscript with inline results. Log everything.”

Executes each step, chains the outputs, renders a reproducible document, and logs every decision to the audit trail.

You don’t have to learn a new language. Just tell it what you need, at whatever level of detail feels right.

Tools built for how scientists actually work

These ship with sciClaw. The agent calls them automatically when your request needs them.

Microsoft Word Review

Real tracked changes & comments in .docx

Ask sciClaw to review your manuscript and it returns a .docx you open in Word, with tracked changes you can accept or reject and comments anchored to specific paragraphs. Exactly like getting a revision back from a co-author.

  • Tracked insertions, deletions, and replacements you review in Word
  • Margin comments tied to specific text, not just line numbers
  • Compare two drafts side-by-side with semantic diff
  • NIH-format Word template included for grant applications

PubMed Search

Full NCBI E-Utilities access

Tell sciClaw what you're looking for and it builds the PubMed query: Boolean operators, MeSH terms, date ranges, publication types. Results export directly to Zotero, EndNote, or Mendeley.

  • Complex queries with MeSH terms, field tags, and filters
  • Walk citation chains forward (cited-by) and backward (references)
  • Find related articles scored by relevance
  • Export as RIS for one-click import into your reference manager
  • Works with your own NCBI API key for faster, higher-volume searches

Frequently asked questions

What do I need to set this up?

A Mac or Linux machine, and an account with an AI provider (your existing ChatGPT subscription works). Install is one Homebrew command. The onboard wizard walks you through the rest. Most people are up and running in a few minutes. You don’t need to touch a terminal after that if you don’t want to.

Where does my data go?

sciClaw runs on your machine. Your files stay local. When you send a message, that conversation goes to whatever AI provider you picked (OpenAI, Anthropic, Google, etc.) using your own API key or subscription. We never see your data. There’s no account with us, no telemetry, nothing phoning home.

Can I actually trust AI output for research?

Same way you’d trust a draft from a new postdoc. You read it critically, push back where it’s wrong, and iterate. The difference here is that sciClaw logs every source, every decision, every revision. Six months later you can retrace exactly how a result came together. If something looks off, just ask. It’s a conversation, not a final answer.

What if I’m not technical?

That’s the whole point. You talk to sciClaw through Telegram or Discord, the same apps you already have open. Type a question, attach a file, get results back. No command line, no code. (Power users can use the CLI if they want, but it’s totally optional.)

What does it cost?

sciClaw itself is free. MIT licensed, open source, no catch. You pay your AI provider directly for the API calls, same as you would for ChatGPT Plus or an Anthropic key. Most researchers spend a few dollars a month.

Does it work on Windows?

Yes. Windows users can run sciClaw through WSL (Windows Subsystem for Linux), which gives you a full Linux environment right inside Windows. Pre-built binaries for macOS (Apple Silicon and Intel), Linux (amd64, arm64), and Windows (amd64) are on the releases page.

Built in the open. Come build with us.

sciClaw is MIT licensed and we genuinely want collaborators. Whether you write Go, work on NLP pipelines, build skills in plain markdown, or just have ideas about how research workflows should work, there’s a place for you.

Core agent

Go codebase. Agent loop, tool orchestration, lifecycle hooks, channel adapters.

Research skills

Write new skills in plain markdown. PubMed, docx-review, and 11 others are just text files you can fork and improve.

Companion tools

Native binaries for document review, PubMed search, PDF handling. Mostly Go with some Python.

View on GitHub Open an issue Join the discussion