data-analyst

An AI-assistant skill (Claude, Codex, or any assistant that reads Agent Skills) that answers questions with official U.S. government data from federal sources (BEA, BLS, FRED, Census, IPUMS, Treasury, OPM, USAspending, CDC, Congress.gov, EPA and more). It picks the sources a question needs, gets the data in any environment, asks only the questions that change the numbers, and returns CSV data plus a short answer, or an audited HTML report or paper when asked.

Each source keeps its own rules (IPUMS microdata is never redistributed; FRED forbids persistent copies; OPM reports disclose redaction coverage; ...). fetch.py --sources lists every source.

Install

Install it as a plugin from the zsu marketplace in the zsu/agent-plugins catalog, which installs it from this GitHub repo. The skill is then /data-analyst:data-analyst, and it also triggers on its own when you ask a question about U.S. government data.

• Claude Code (terminal):

  /plugin marketplace add zsu/agent-plugins
  /plugin install data-analyst@zsu
  

  or from a shell: claude plugin marketplace add zsu/agent-plugins, then claude plugin install data-analyst@zsu.
• Claude Desktop (Code tab): click +, choose Add plugins, add the marketplace zsu/agent-plugins, then install data-analyst.
• Cowork and claude.ai (web): Customize > Plugins > Add > Add marketplace, enter zsu/agent-plugins, then install data-analyst. Or run python tools/package_skill.py --plugin and use Add > Upload plugin with dist/data-analyst-plugin.zip. Keyless sources work when the organization's network settings allow their hosts; otherwise the skill hands over a manual-download bundle and ingests the uploads.
• Skill only on claude.ai: python tools/package_skill.py, then upload dist/data-analyst.zip under Settings > Capabilities > Skills.
• A whole team's repo: commit this to the project's .claude/settings.json; collaborators are prompted to install it when they trust the folder:

  {
    "extraKnownMarketplaces": {"zsu": {"source": {"source": "github", "repo": "zsu/agent-plugins"}}},
    "enabledPlugins": {"data-analyst@zsu": true}
  }

• Codex (CLI and app): codex plugin marketplace add zsu/agent-plugins, then install data-analyst from the Plugins directory. The skill triggers on its own for U.S. government data questions. Codex's sandbox may block network access; the skill says so and how to allow it.
• Any other assistant that reads Agent Skills: copy plugins/data-analyst/skills/data-analyst/ (or unzip dist/data-analyst.zip from python tools/package_skill.py) into that assistant's skills folder.
• Develop without installing: from a clone, claude --plugin-dir plugins/data-analyst.

The skill never writes into its own folder (the plugin cache): outputs go to out/ in your working directory (or $DATA_ANALYST_OUT) and downloads are cached in ~/.cache/data-analyst.

API keys (optional for most sources)

Run the setup script for a source in your own terminal; it asks for the key with hidden input and stores it in ~/.config/data-analyst/keys.json (or $DATA_ANALYST_CONFIG_DIR), which the skill reads directly under any assistant. If the key variable (e.g. FRED_API_KEY) is already set in your environment, the script stores nothing: the skill uses it (--force / -Force stores one anyway). Keys stored earlier in Claude Code's settings.json keep working (--claude / -Claude still writes there). When the skill needs a key it prints the exact command for your install; from a clone, run:

bash plugins/data-analyst/skills/data-analyst/setup.sh fred

powershell -NoProfile -ExecutionPolicy Bypass -File plugins/data-analyst/skills/data-analyst/setup.ps1 -Source fred

python plugins/data-analyst/skills/data-analyst/scripts/fetch.py --sources lists every source and its key variable.

How it is built

Path	What
plugins/data-analyst/skills/data-analyst/SKILL.md	the workflow the model follows
.../scripts/fetch.py	one command line for every source, plans of several sources, routes
.../scripts/core/	keys, environment, HTTP client, routes, cache, packaging, manual bundles, policy
.../scripts/analysis/, .../scripts/report/	Facts registry, combining helpers; report builder and audit
.../sources/<id>/	one pack per source: policy.json (enforced rules), source.md, endpoints, tests
docs/pack_contract.md	how to add a source
tests/	offline tests (core, report chain, sandbox) and each pack's tests and fixtures in tests/sources/<id>/
tools/	routing generator and upload packager

Tests: python -I tests/run_all.py.