ai-anthropology-toolkit

PyPI

Computational tools for anthropological and qualitative research: data collection and analysis as an MCP server and Python package

Versions collected1

Direct dependencies22From release 3.9.0

Last observed21 Oct, 09:55 UTC

README

Version 3.9.0 documentation

AI Anthropology Toolkit

A suite of AI anthropology tools for qualitative research

Matt Artz | GitHub | ORCID

Overview

The AI Anthropology Toolkit provides computational tools for anthropological and qualitative research. Every component is grounded in the conventions, debates, and craft knowledge of anthropology and cognate qualitative social sciences. Epistemic stance (interpretivist, critical, STS, feminist, applied, etc.) is treated as a first-class design parameter that shapes methods, writing, and analysis.

The toolkit ships as plugins for Claude Code and Codex, an MCP server that lets Claude, Codex, or Gemini run the full pipeline — from data collection through coding, thematic analysis, and networks of codes — directly, and, for researchers who want to work closer to the code, standalone notebooks for data collection and qualitative analysis.

What is AI Anthropology?

AI Anthropology is the integrated practice of studying, using, and shaping AI (Artz 2026) — an iterative lifecycle of anthropology of, by, and for AI (Artz 2026):

Anthropology of AI — studying AI systems as cultural phenomena: how they reflect, reshape, and redistribute human culture and power
Anthropology by AI — using AI and computational methods to extend ethnographic research while preserving interpretive judgment
Anthropology for AI — bringing anthropological insight to the design, building, and governance of AI systems

This toolkit primarily operationalizes anthropology by AI — scaling qualitative research while keeping the researcher's interpretive authority at the center — and is itself an act of anthropology for AI: purpose-built research infrastructure designed with anthropological sensibilities. Its governing design philosophy is friction by design: divergence between machine readings is sustained as data rather than resolved as noise, and the interaction itself withholds production wherever a judgment is not the machine's to make (see skills/DESIGN.md). The toolkit is presented as a demonstration of multi-agent ethnography in A Call for an AI Anthropology (General Anthropology) and Multi-Agent Ethnography: Post-Conventional Anthropological Practice Through Human-AI Collaboration (Anthropological Forum).

Skills

Research skills in the portable SKILL.md format that activate automatically based on context. In Claude Code, the AI Anthropology Toolkit plugin installs all of them at once; other coding agents can install any skill by copying its folder (see Using the Skills in Other Agents below).

Skill	Description
research-question	Helps you turn a research interest into a clear, answerable research question. Walks through the five parts a good question needs and checks it against a rubric.
literature-review	Helps you plan and write a literature review, whether narrative, scoping, or systematic. Guides your search strategy, logs what you screened and why, and helps you build the bibliography or framework you need.
methodology-selection	Helps you choose and justify research methods that fit your question and your stance. Checks that your methods actually match what you are trying to find out, including when you need more than one.
research-plan	Helps you write a standalone research plan. Covers every section, from the problem statement through feasibility.
irb-protocol	Helps you write your IRB protocol. Covers risk assessment and the specific ethics of digital ethnography, section by section.
informed-consent	Helps you design your informed consent process. Covers different consent modes, like written, verbal, layered, or community-based, and how to adapt them to your cultural context.
grant-proposal	Helps you write a grant proposal for a specific funder. Covers NSF CA-DDRIG, Wenner-Gren, Fulbright, ERC, SSHRC, and Wellcome, each with its own guidance.
dissertation-prospectus	Helps you write your dissertation prospectus, section by section. Works whether your prospectus is short or long.
fieldwork-methods	Helps you design your fieldwork instruments. Covers interview guides, observation protocols, sampling strategy, and how you will manage your data.
qualitative-analysis	Helps you analyze your qualitative data. Covers building a codebook, coding it, and pulling out themes, including comparing across more than one analytical lens.
ethnographic-generalization	Helps you figure out what your confirmed findings actually generalize to. Walks through the kind of claim you can make, builds its warrant, and states its scope and how confident you can be.
digital-computational-methods	Helps you design digital or computational research methods. Covers digital ethnography, platform ethics, large-scale text analysis, and working alongside AI tools.
paper-planning	Helps you work out what your paper actually argues before you draft it. Extracts your claim, positions it against existing work, and sequences your argument through questions rather than writing.
research-writing	Helps you write your article or dissertation chapter. Covers its structure, ethnographic craft, and the conventions of your subfield or journal.
abstract-writing	Helps you write the abstract, title, and keywords for a journal or thesis submission. Builds the abstract from your manuscript's own claim, cuts it to the word limit and tells you what the cut cost, checks that it does not promise more than the manuscript delivers or name a site your anonymization protects, and picks keywords that index what the title leaves out.
academic-review	Helps you write peer reviews and respond to them. Covers writing the review itself, rebuttal letters, and your revision strategy.
rival-interpretations	Helps you test a claim against rival readings before a reviewer does. Argues your material from three other analytical positions, separates what they agree on from what stays genuinely open, and leaves you a record for your methods section.
manuscript-markup	Helps you work through a manuscript that came back marked up with comments. Reads each comment in place, sorts them, works through them, and drafts your reply letter.
review-anonymization	Helps you prepare the anonymized copy of your manuscript for double-blind peer review. Keeps your full version as the master, swaps author names and self-citations for placeholders without rewording anything, finds the leaks a placeholder cannot fix, such as your fieldsite and fieldwork history or a grant number, clears the hidden metadata in your Word file, and searches the finished copy for anything that still names you. Never touches your participants' pseudonyms.
proof-review	Helps you check a publisher's typeset proof against the manuscript you submitted. Compares them word by word, checks that your pseudonyms, quoted speech, and citations survived production, and writes the correction list you send back.
conference-materials	Helps you prepare for a conference. Covers abstracts, slide decks, posters, speaker notes, and how you will deliver the talk.
public-engagement	Helps you write for a general audience. Covers op-eds, blog posts, policy briefs, and community reports.
job-materials	Helps you prepare your job market materials. Covers your CV, cover letter, job talk, and overall application strategy.
career-statements	Helps you write your career statements. Covers research, teaching, and diversity statements, and tenure narratives.
teaching-materials	Helps you build your course materials. Covers syllabi, lesson plans, assignments, rubrics, and discussion guides.
applied-practice	Helps you write client-facing deliverables. Covers statements of work, stakeholder readouts, insight synthesis, and workshop materials.
repeated-work	Helps you decide whether work you keep repeating is actually worth turning into a tool. Only if it is, helps you figure out what kind of tool it should be.
tool-building	Helps you build your own research tool, like a scraper, MCP server, skill, or agent. Works out a specification with you before writing any code.

Using the Skills in Other Agents

Each skill is a self-contained folder (SKILL.md plus a references/ directory) in the portable format that most 2026 coding agents read. To install one outside Claude Code, clone the repository and copy the skill folder — plus skills/DESIGN.md, the shared analytical-lens reference the skills consult — into your agent's skills directory:

git clone https://github.com/MattArtzAnthro/AI-Anthropology-Toolkit.git
cp -r AI-Anthropology-Toolkit/skills/qualitative-analysis AI-Anthropology-Toolkit/skills/DESIGN.md ~/.codex/skills/

Agent	Skills directory
Claude Code	`~/.claude/skills/` (or install the plugin — all 28 at once)
OpenAI Codex CLI	`~/.codex/skills/`
Cursor	`~/.cursor/skills/`
GitHub Copilot / VS Code	`~/.copilot/skills/`
Shared project-level	`.agents/skills/` in your repository

The skills pair naturally with the MCP server (below): descriptions route the request, the server runs the pipeline.

Agents

Autonomous Claude Code subagents that orchestrate across multiple skills for complex, multi-step tasks.

Agent	Description
research-design	Orchestrates question, methodology, and plan skills for end-to-end research design
ethics-reviewer	Reviews research designs for ethics issues, drafts protocols and consent documents
proposal-advisor	Translates research designs into persuasive funder-specific narratives
fieldwork-advisor	Designs instruments tailored to specific research questions and fieldwork contexts
analysis-advisor	Guides qualitative coding, codebook development, and thematic analysis
writing-advisor	Guides argument planning, article/chapter writing, and R&R management
dissemination-advisor	Handles register translation between academic and public audiences
career-advisor	Coordinates application packages and course design
network-analyst	Builds and reads networks of coded material (codes that co-occur, speakers to codes, lens agreement) with the MCP network tools; no Gephi required
tool-builder	Builds a research instrument, or a skill, agent, or MCP tool for this toolkit. The only agent here that writes files

Commands

Command	Description
`/ai-anthropology:new-project`	Scaffold a new research project through guided phases
`/ai-anthropology:build-tool`	Build a research instrument, specification first
`/ai-anthropology:test-claim`	Test one interpretive claim against rival readings, and record what stays open
`/ai-anthropology:skills`	List the toolkit's skills, agents, and commands

MCP Server

The toolkit also ships as a Python package (ai-anthropology-toolkit on PyPI) with an MCP server, so Claude (and other MCP clients) can drive the full research pipeline conversationally: data collection (OpenAlex, CrossRef, PubMed, Google Scholar, Google Trends, Google News, Google Patents, Books Ngram, YouTube search and transcripts, podcast RSS), citation formatting (a bare DOI rendered in any CSL style, including the anthropology journal styles), and analysis (transcript chunking, lens-configured codebook generation, qualitative coding with per-code validation, thematic analysis, and cross-lens comparison).

Installing either bundled plugin registers the server automatically. It can also be registered directly in another MCP-capable agent:

Claude Code

claude mcp add ai-anthropology -- uvx --from "ai-anthropology-toolkit[data,api]==3.9.0" ai-anthro-mcp

OpenAI Codex CLI

codex mcp add ai-anthropology -- uvx --from "ai-anthropology-toolkit[data,api]==3.9.0" ai-anthro-mcp

Google Gemini CLI

gemini mcp add -s user ai-anthropology uvx -- --from "ai-anthropology-toolkit[data,api]==3.9.0" ai-anthro-mcp

The server also builds and reads networks of coded material with no Gephi installed: build_network (codes that co-occur, speakers to codes, lenses by agreement), analyze_network (layout, communities, centralities, a visual-QA reading), view_network (an interactive map in chat where the host renders MCP Apps), and export_network (GEXF for Gephi, where Gephi AI takes over).

The server is model-agnostic. By default, whichever model is orchestrating — Claude, GPT, or Gemini — performs each interpretive step itself through validated work packets (delegated mode): the analysis runs on your model, the methodology and validation run on the server, and every coding decision stays visible to the researcher. To have the server call Claude itself instead (api mode, billed to your Anthropic account), set both ANTHROPIC_API_KEY and AAT_MODE=api in the server's environment, for example in the env block of its MCP configuration. A key on its own never switches api mode on, and the orchestrating model cannot switch it on for you.

Coding Agents & Sandboxes

AI coding agents — Claude Code and Claude Desktop's Cowork, OpenAI Codex CLI, Gemini CLI — often run in sandboxed environments where MCP connectors are unavailable but pip and Python still work. The toolkit degrades gracefully across three tiers:

MCP tools available → use them; the full pipeline runs natively.
Code execution only → install the package and call the Python API directly:
```
pip install "ai-anthropology-toolkit[data]"
python -m ai_anthro_toolkit.doctor
```
The doctor (python -m ai_anthro_toolkit.doctor, also installed as ai-anthro-doctor) probes every data source from the current network and reports which are reachable. Sandbox network policies typically allow the scholarly APIs (OpenAlex, CrossRef, PubMed) and block the Google/YouTube scraping endpoints — collect what is reachable and route each blocked source to local execution or its Colab notebook (the doctor prints the link). The collector functions live in ai_anthro_toolkit.datasources; transcript chunking (ai_anthro_toolkit.chunking, fully local) and the 42-lens registry (ai_anthro_toolkit.lenses) work in any environment. Agent-facing instructions for this fallback chain ship in this repository as AGENTS.md and GEMINI.md.
No code execution either → every capability runs in the browser through the Colab notebooks below.

Checking an artifact

ai-anthro-check path/to/codebook.json      # or: python -m ai_anthro_toolkit.checks <path>

Runs the standing checks over a codebook or a coded dataset and reports what they found. Most of these run on their own, without being asked, whenever the toolkit produces one of those artifacts; the command exists so you can re-run them later, on an artifact you were handed, without the pipeline that made it.

A check that fires is a question rather than a verdict. It names a commitment the artifact implies — that codes were meant to be mutually exclusive, that every code in the book was meant to earn its place, that the codebook did not move mid-pass — and only you can say whether that commitment is yours. A check that could not run is reported as unrun, never as passed, and a run in which nothing capable of surprising you executed says so.

When you build an instrument through tool-building, it writes instrument-checks.json beside your data from the commitments you settled while specifying it — what an empty result means, whether a repeated identifier is an error, whether a field is required. ai-anthro-check picks that file up automatically, so the checks are about your artifact rather than only the toolkit's.

The file says what it cannot do. A commitment that cannot be checked from the data alone — whether source order was preserved, for instance, which only the source can settle — is listed as unenforceable with the reason. If you settled five commitments and three became checks, it tells you which two did not and why.

One flag matters: --distinct-codes if you hold codes to be mutually exclusive, --overlapping-codes if you keep overlapping codes deliberately, as grounded theory and several interpretive traditions do. Left unsaid, the check that depends on it does not run, because that is a methodological commitment and not the toolkit's to assume.

Notebooks

Standalone notebooks for computational qualitative research. Most run directly in Google Colab; notebooks marked Local should be run on your own machine (see Running Locally below).

Data Collection

Gather scholarly, public, and media data through official APIs and structured collectors.

Notebook	Run	Description
Academic Literature Explorer		Search 250M+ scholarly works across all disciplines via OpenAlex with citation counts and open access detection
CrossRef Reference Verifier		Verify reference lists against canonical CrossRef metadata — DOI resolution, text-vs-record comparison, retraction flags — plus journal and author queries
PubMed Literature Harvester		Search PubMed and enrich results with metadata from CrossRef, OpenAlex, and Semantic Scholar
World Bank Data Explorer		Retrieve development indicators from the World Bank API — 16,000+ series across 200+ economies, no API key required
UN Data Explorer		Retrieve UN development data — SDG indicators (no key) and UNDP Human Development indices (free API key)
BLS Labor Statistics Explorer		Retrieve US labor and economic series from the Bureau of Labor Statistics — unemployment, prices, wages — keyless
Google Books Ngram Explorer		Analyze historical word frequency patterns across Google Books corpora (1800-2022) with visualization and export
Google Trends Explorer		Retrieve and visualize Google Trends data with multi-term comparison, regional breakdowns, and related queries
Google News Explorer		Search Google News by keyword, time period, and country with quick or extended date-range modes
Google Scholar Explorer	Local	Search Google Scholar for publications with year filtering, citation counts, and structured export
Google Patents Explorer		Search Google Patents for patent metadata including titles, inventors, assignees, and filing dates
YouTube Video Search		Search YouTube and export video metadata including titles, channels, views, and durations
YouTube Transcript Fetcher		Fetch YouTube video transcripts with language selection, segment chunking, and multiple export formats
Podcast RSS Explorer		Pull episode metadata from any podcast RSS feed with titles, dates, durations, and structured export

Qualitative Analysis Pipeline

Four notebooks cover the path from recording to themes — transcribe, segment, build a codebook from your source literature, then code and construct themes. The transcript chunker and the codebook builder both feed coding and thematic analysis. Each also works standalone.

Notebook	Run	Description
Audio Transcription with Whisper		Transcribe audio and video recordings locally with Whisper — timestamped transcripts, optional speaker diarization, and Chunker-ready export
Interview Transcript Semantic Chunker		Segment interview transcripts into semantically coherent chunks with speaker-aware processing and coherence scoring — fully local, no API key required
Qualitative Codebook Builder		Build qualitative codebooks from source literature with AI-assisted code generation, validation, and structured export
Coding and Thematic Analysis		Apply codes to qualitative data and build themes using deductive, inductive, or hybrid approaches, with multi-lens parallel analysis and cross-lens comparison

Multimodal & Visual Analysis

Work with images, audio, video, and mixed-media collections through Gemini's multimodal models (free API key from Google AI Studio).

Notebook	Run	Description
Multimodal Embedding Explorer		Embed text, images, audio, video, and PDFs into one vector space — similarity, clustering, UMAP maps, and cross-modal search
Visual Analysis & Image Annotation		Annotate field photos with Gemini vision — ethnographic descriptions, structured tags, and categorization

Computational Text Analysis

Standalone notebooks for analyzing text corpora at scale.

Notebook	Run	Description
Text Network Analysis		Build co-occurrence networks from text with community detection, centrality metrics, and interactive visualization
Topic Modeling (BERTopic)		Discover topics in text collections using transformer-based clustering with interactive visualizations and zero-shot mode
Named Entity Recognition (GLiNER2)		Extract people, places, organizations, concepts, and custom entity types from text using zero-shot NER

Releasing

Maintainer checklist for cutting a release, covering the three version tracks, the upload-before-push ordering, and how to verify both dependency resolution and live-host discovery: RELEASING.md.

Companion Plugins

Where a specialized tool already owns a capability, the toolkit hands off to it instead of duplicating it here.

Plugin	Capability	Install
gephi-network-analysis	Network analysis in live Gephi Desktop — text-network construction, layouts, centrality and community metrics, structural claim verification	`claude plugin marketplace add MattArtzAnthro/gephi-ai`, then `claude plugin install gephi-network-analysis@gephi-ai`

With the companion installed, the digital-computational-methods skill hands network-analysis execution to it. Without it, the skill falls back to the Text Network Analysis notebook and a GEXF export for manual work in Gephi.

Getting Started

Claude Code Plugin

Install the plugin in Claude Code:

/plugin marketplace add MattArtzAnthro/AI-Anthropology-Toolkit
/plugin install ai-anthropology@ai-anthropology

Skills activate automatically when Claude detects relevant context. Agents handle multi-step tasks across skills. Commands are invoked with slash syntax.

Codex Plugin

Clone this repository, then add it as a local marketplace and install the bundle:

codex plugin marketplace add /absolute/path/to/AI-Anthropology-Toolkit
codex plugin add ai-anthropology@personal

Start a new Codex task after installation. The plugin includes all 28 portable research skills, a Codex-specific router, and the same exactly pinned 34-tool MCP server registration as the repository root.

Notebooks (Colab)

Click any Open in Colab badge in the Notebooks section to run a notebook directly in your browser. Each notebook handles its own dependencies — no local installation needed.

Running Locally

Some notebooks (marked Local in the table) need to be run on your own machine. This requires Python and Jupyter.

If you already have Anaconda/Miniconda installed:

pip install scholarly
jupyter notebook

Then open the notebook file from the Jupyter file browser.

If you need to install Jupyter from scratch:

pip install jupyter scholarly
jupyter notebook

Notebooks that run locally will install any other dependencies they need automatically when you run the first cell.

License

This repository — notebooks, Python package, MCP server, plugin content, and documentation — is licensed under the PolyForm Noncommercial License 1.0.0: free for noncommercial use, including research, education, and work by nonprofit and government research organizations, with attribution to Matt Artz appreciated. For commercial licensing, contact Matt Artz.

Citation

If you use this toolkit in your academic research, please cite:

Artz, Matt. 2025. AI Anthropology Toolkit. Software. Zenodo. https://doi.org/10.5281/zenodo.16728812

Research Behind the Toolkit

The toolkit operationalizes a published research program on AI Anthropology:

Artz, Matt. 2026. "A Call for an AI Anthropology." General Anthropology 33(1): 23–28. https://doi.org/10.1111/gena.70007.

Artz, Matt. 2026. "Multi-Agent Ethnography: Post-Conventional Anthropological Practice Through Human-AI Collaboration." Anthropological Forum. https://doi.org/10.1080/00664677.2026.2614501.

Artz, Matt. 2026. "Artificial Intelligence: The AI Anthropology Lifecycle (of, by, for AI)." In Practicing Digital Ethnography, edited by Devin Proctor. Routledge. https://doi.org/10.4324/9781032672663-29.

Koycheva, Lora, Angela K. VandenBroek, and Matt Artz, eds. 2026. Anthropology and AI. New York: Routledge. https://doi.org/10.4324/9781003532750.

Artz, Matt. 2023. From Machine Learning to Machine Knowing: A Digital Anthropology Approach for the Machine Interpretation of Cultures. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000384902.

Artz, Matt. 2023. "Ten Predictions for AI and the Future of Anthropology." Anthropology News, May 8. https://doi.org/10.1111/AN.1605.

Artz, Matt. Forthcoming. "AI Anthropology: The Future of Applied Anthropological Practice." In Routledge Handbook of Applied Anthropology, edited by Christina Wasson, Edward B. Liebow, Karine L. Narahara, Ndukuyakhe Ndlovu, and Alaka Wali. New York: Routledge.

Downloads & popularity

Awaiting observations

No metric observations yet

Download counts and popularity will appear here as the collector receives them.

Versions 1

Collected releases

Version	Published	Direct dependencies	Open version
3.9.0	12 Oct 2026	22

Package advisories

No matching advisories collected

This describes our collected data; it is not a security assessment.

Repository

PENDING

mattartzanthro/ai-anthropology-toolkit

Shared repository data · GitHub

License detected in repository: Not collected

Metadata history

Meaningful changes

Current collected state Current

Observed since 21 Oct, 09:55 UTC

No superseded metadata states collected. Unchanged observations do not create new revisions.

Record identity

package:69d6bbd5cf28d2d4d3fd37f76fa7113deb9dedd16485c4e0835c52d4d8ae7ab1