open-apis

Autonomously curated catalog of free, open Norwegian APIs — built for AI code agents.

What this is

A machine-readable catalog of Norwegian public APIs (weather, transport, business register, statistics, geodata, …) with metadata that AI coding agents actually need: when to use it, what to chain it with, sample requests, and citation-grounded descriptions.

The catalog is self-maintaining: a daily pipeline harvests Felles datakatalog, proposes new entries via curator subagents, verifies citations, runs live smoke tests, and opens PRs.

Status

Bootstrap — the schema is in place, the first hand-curated entry (met.locationforecast) validates end-to-end, and the harvester pulls live FDK data.

Browse the catalog

A generated landing page lives at derived/index.md — human-readable index of every entry grouped by status, with views by category and by provider. Machine consumers should read derived/index.json.

These artifacts are rebuilt by uv run open-apis index from the YAML under catalog/. They are derived data — never edit by hand.

Quick start

uv sync
uv run open-apis schema      # regenerate catalog/schema.json
uv run open-apis validate catalog/by-provider/met/locationforecast.yaml
uv run open-apis smoke    catalog/by-provider/met/locationforecast.yaml
uv run open-apis harvest dataservices 2   # 200 entries from FDK
uv run open-apis index       # rebuild derived/ landing + index.json

Architecture

See CLAUDE.md for the autonomy model. Short version:

• Source of truth: pipeline/models.py (Pydantic) → catalog/schema.json
• Catalog: catalog/by-provider/<provider>/<api>.yaml (canonical) and catalog/staging/... (quarantine)
• Raw artifacts: raw/<provider>/<date>/... (immutable, citations point here)
• Pipeline: pipeline/{harvest_fdk,validate,smoke,citations}.py
• Agent infrastructure: .claude/{agents,skills,hooks,rules}/

Autonomy split: deterministic CI + local LLM curation

• GitHub Actions runs the pure-Python validators and smoke tests. No LLM. No API key. Free.
• LLM curation runs on your machine via a Claude Code scheduled task (~/.claude/scheduled-tasks/daily-harvest/). It uses your existing Claude Code subscription — no Anthropic API billing.

This split is intentional: the deterministic gates (schema, citation substring match, smoke) live in CI where they protect the main branch from any contributor; the agentic curation happens locally where it's free under your subscription.

Trust model

Every textual field in a catalog entry has a source_quote that must literally appear in the raw artifact it cites. The validator does this check with a string substring match — not another LLM. This is the deterministic anti-hallucination gate.

A separate citation-reviewer subagent (read-only, no write tools) independently verifies each entry. A Stop hook prevents Claude Code from terminating with unreviewed changes to catalog/.

License

• Code: MIT
• Catalog metadata: NLOD-2.0 (matches the upstream Felles datakatalog license)

Contributing

This is a single-developer + agent project. PRs from humans are welcome but the daily harvester does the bulk of the work.