Icelandic Data

Data toolkit for Icelandic public data. The .agents/skills/ files document each data source — API endpoints, series codes, encoding quirks, classification changes — and the scripts/ directory has Python scripts that fetch, clean, and transform the data. It’s harness-agnostic: any agent that can read the skills and run the scripts they reference (Claude Code, Codex, and others) works.

“I spent a good chunk of my career, including an entire startup, DataMarket, making public data accessible. This project cracks the problem from the angle I now believe is the right one: don’t build another portal, make the data agent-ready. Ask a question in plain language and watch it pull together answers from dozens of official sources. I’ve been showing it to people in government as a direct demonstration of what such an approach can deliver.”

  Hjálmar Gíslason
  Founder & CEO of GRID
  Previously founder of DataMarket (acquired by Qlik)

The data sources badge reports upstream health, not this repo’s code: a source only turns it red after 7 straight days down or returning the wrong shape. Anything shorter is weather — these are public APIs and they blink. Details in AGENTS.md; raw history is on the health-history branch.

Not a portable skill library. The skills reference co-located scripts and work as a unit. Get the repo, run setup, and use an agent to research questions, join data sources, or produce outputs — a gist, a CSV, an HTML report, whatever fits. It is plain Python and runs on macOS, Linux and Windows.

Structure

.agents/skills/{source}/SKILL.md → Data source docs (API, endpoints, caveats)
scripts/{source}.py              → Fetch + transform scripts
tests/health/test_{source}.py    → Upstream health probes (marker: health)
data/raw/{source}/               → Raw downloads
data/processed/                  → Cleaned datasets

Skills follow the agentskills.io open standard. .agents/skills/ is the real location (read natively by Codex); .claude/skills symlinks to it for Claude Code, and CLAUDE.md symlinks to AGENTS.md. One set of files, both agents.

Data sources

Currently 56 skills covering national statistics, government dashboards (mælaborð), regulatory filings, and utility APIs. Each skill’s frontmatter description is its index entry — run ls .agents/skills/ to enumerate them, and see AGENTS.md for the quick-commands reference.

The light in each row is that source’s current verdict, refreshed daily and finer-grained than the badge above (which waits a full week before going red): healthy · flaky (failed but recovered — these APIs blink) · dead, or the skill no longer matches the source · not enough observations yet. A · means there is nothing upstream to probe. Per-source uptime and last error are in the workflow summary.

Statistics & macroeconomic

Government dashboards (mælaborð)

Systematic coverage of public dashboards published under Iceland’s data-access law.

Business & markets

Property, planning, addresses

Transport & mobility

Environment & geography

Energy & fisheries

Personal finance & rates

Legal & civic

Setup

You need a terminal once, for setup. After that, an agent does the typing.

1. Get the repo. git clone https://github.com/jokull/icelandic-data — or, without git, use Code → Download ZIP on GitHub and unzip it. (The Code tab in Claude Desktop needs Git for Windows installed on Windows even for a ZIP checkout; Claude Code CLI and Codex do not.)

2. Run setup. It installs uv if you don’t have it, which then downloads Python and every dependency. Nothing else is required — no Homebrew, no system Python.

   ./setup.sh                                          # macOS / Linux
   powershell -ExecutionPolicy Bypass -File .\setup.ps1  # Windows
   

   Setup ends by running scripts/setup_check.py, which reports what works and repairs the .claude/skills link if git or the ZIP download flattened it. (On Windows that repair is a junction, and git status will show .claude/skills as changed from then on — expected, do not revert it.)

3. Open the folder in your agent. Claude Code, Codex, the Code tab in Claude Desktop, or ChatGPT Desktop (Codex) all read the skills from the repo. Ask a question such as “What is the median price per m² of apartments in Reykjavík since 2020?” Codex asks whether to trust the folder the first time — say yes. That loads .codex/config.toml, which lets the sandbox reach the Icelandic APIs; without it every fetch fails with a DNS error.

That is Tier 0 and covers 41 of the 56 skills. The rest need one more opt-in install; each skill says which at the top of its SKILL.md:

scripts/setup_check.py prints which tiers are installed.

What a first run looks like

Measured on a fresh macOS user account and a bare Ubuntu container, both with only curl on PATH: setup.sh finishes in 20–30 s, the environment is about 500 MB, and the maps tier adds a couple of seconds. Where newcomers can stall, and the fix:

Every command in the docs is of the form uv run python scripts/.py ... and works the same in bash, zsh, fish and PowerShell. SQL against the processed files goes through uv run python scripts/sql.py "SELECT ...", so no DuckDB binary is needed.

Usage

# Process data
uv run python scripts/sedlabanki.py
uv run python scripts/fuel.py

# Query with DuckDB
uv run python scripts/sql.py "SELECT * FROM 'data/processed/fuel_prices_daily.csv' LIMIT 10"

# Company financials pipeline
uv run python scripts/financials.py company  --year 2024

# Scrape a government dashboard
uv run python scripts/landlaeknir.py list
uv run python scripts/landlaeknir.py fetch --slug mortis

# Build the indicator catalog for visar.hagstofa.is
uv run python scripts/velsaeldarvisar.py fetch

See AGENTS.md for the full command catalog across every skill.

Tests

uv run pytest -m "not slow"   # fast unit tests (~0.5s) — what PR CI runs
uv run pytest -m slow         # network + Playwright tests (several minutes)
uv run pytest -m health       # upstream health probes
uv run pytest -m health -k hagstofan   # probe one source

Health probes check that each upstream source still serves the smallest contract its script depends on. They run daily in CI; browser-based probes are manual-dispatch only. See .github/workflows/source-health.yml.

Adding a new data source

See the new-data-source skill for the methodology — discovery, probing, skill authoring, script conventions, testing, and visualization.