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.
.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.
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.
Systematic coverage of public dashboards published under Iceland's data-access law.
| Source | Description | |
|---|---|---|
| Samgöngustofa | Vehicle registrations by make, fuel type, location — Power BI | |
| car (island.is) | Per-vehicle lookup by plate/VIN via public GraphQL API |
| Source | Description | |
|---|---|---|
| Laun (payday.is) | Take-home salary calculator with tax/pension breakdown | |
| Gengi (Borgun) | Currency exchange rates (card rates, not interbank) |
You need a terminal once, for setup. After that, an agent does the typing.
-
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.) -
Run setup. It installs
uvif 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/skillslink if git or the ZIP download flattened it. (On Windows that repair is a junction, andgit statuswill show.claude/skillsas changed from then on — expected, do not revert it.) -
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:
| Tier | Install | Unlocks |
|---|---|---|
| 0 core | uv sync |
statistics, dashboards, registers, finance, polls, weather — plain Python |
| 1 maps | uv sync --group maps |
map rendering and rasters (kortagerð, LMI, NÍ habitats, traffic/cattle maps) |
| 2 browser | uv run playwright install chromium |
Power BI / Tableau / JS-rendered scrapers |
| 3 pdf | uv sync --group pdf |
Docling AI table extraction (pulls in torch, ~2 GB) |
| 4 restricted | — | needs an Icelandic IP or private credentials; documented, not portable |
scripts/setup_check.py prints which tiers are installed.
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:
| Stops at | Why | Fix |
|---|---|---|
| Claude Desktop Code tab on Windows | needs Git for Windows | install Git, restart the app |
PowerShell refuses setup.ps1 |
execution policy | run it as shown above, with -ExecutionPolicy Bypass |
uv: command not found right after install |
new PATH not in the current shell | open a new terminal, or re-run setup.sh |
| Codex: every fetch fails with a DNS error | sandbox has no network until the folder is trusted | trust the folder when asked |
| Power BI / Tableau skill fails to start a browser | Tier 2 not installed | uv run playwright install chromium |
| A map script exits asking for the maps tier | Tier 1 not installed | uv sync --group maps |
| Samgöngustofa times out | geo-fenced to Icelandic IPs | run from Iceland or skip |
On Windows setup.ps1 |
||
sets PYTHONUTF8=1 for your user account so Icelandic characters print correctly. |
Every command in the docs is of the form uv run python scripts/<name>.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.
# 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 <kennitala> --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 fetchSee AGENTS.md for the full command catalog across every skill.
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 sourceHealth 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.
See the new-data-source skill for
the methodology — discovery, probing, skill authoring, script conventions,
testing, and visualization.

