okfgen v0.1.3 — ISO 8601 timestamps (OKF v0.2 revision)

Keeps okfgen faithful to the latest OKF v0.2 spec.

Google revised the OKF v0.2 specification in place (same version number) to require that every timestamp-valued key be an ISO 8601 datetime with an explicit UTC offset, not a bare YYYY-MM-DD date. okfgen's built-in spec-drift watcher caught it, and this release brings the output in line.

What changed

• last_modified in sources is now emitted as a full ISO 8601 UTC datetime (e.g. 2026-04-15T00:00:00Z) instead of a date.
• The Socrata and git/local producers pass full timestamps through.
• generated.at already carried a UTC offset — unchanged.

Conformance rules are unchanged, so existing bundles still validate; this just makes okfgen's provenance timestamps match the revised spec exactly.

Install / upgrade

pip install -U okfgen

Full Changelog: v0.1.2...v0.1.3

okfgen v0.1.2 — turn any source into an Agent Skill

okfgen can now turn any source into an Agent Skill.

Highlights

New: okfgen skill — turn a repo, database, docs site, open-data portal, or an existing OKF bundle into an Agent Skill folder:

okfgen skill ./my-repo -o my-skill

It produces a SKILL.md with:

• a name and a triggering description (what it covers and when to use it, with real keywords pulled from the source),
• lean instructions (how to use the reference, respect provenance, don't invent facts),

and the OKF bundle alongside under reference/ — progressive disclosure, so an agent loads detail only when it needs it. Deterministic by default; --llm sharpens the description.

Also available as the okfgen_skill MCP tool, so Claude/Cursor can say "make a skill from this repo" and get one.

Install / upgrade

pip install -U okfgen        # or: uvx okfgen skill ./my-repo -o my-skill

Links

• 🎨 Live demo gallery: https://bushans.github.io/okfgen/
• 📦 PyPI: https://pypi.org/project/okfgen/

Full Changelog: v0.1.1...v0.1.2

okfgen v0.1.1 — OKF v0.2 alignment + provenance

okfgen now tracks OKF v0.2 — with provenance and real credibility signals.

Google bumped the Open Knowledge Format from v0.1 to v0.2 (a backward-compatible minor bump). okfgen's built-in spec-drift watcher caught it, and this release brings okfgen in line — and puts the new provenance model to work.

Highlights

OKF v0.2 alignment

• Bundles now declare okf_version: "0.2".
• Each concept records how it was produced: generated: { by: okfgen/…, at: … } (the v0.2 replacement for the retired timestamp field).
• Validator, consumer, and conventional headings updated to v0.2; existing v0.1 bundles still validate and load.

Provenance with real credibility signals (sources)
Every concept now records what it derives from in the v0.2 sources family — including the objective credibility signals the source actually exposes:

• CKAN: publisher + last-modified per dataset/resource.
• Socrata: publisher, real view counts (usage_count), and last-updated — e.g. NYC 311 comes through with usage_count: 1281089.
• git / local: per-file and per-module last_modified (mtime).
• schema / BigQuery / Firebase: resource + last-modified where available.

Install / upgrade

pip install -U okfgen        # or: uvx okfgen generate .

Links

• 🎨 Live demo gallery: https://bushans.github.io/okfgen/
• 📦 PyPI: https://pypi.org/project/okfgen/

Full Changelog: v0.1.0...v0.1.1

okfgen v0.1.0 — make your systems agent-readable in one command

okfgen turns any repo, database, docs site, or open-data portal into a portable, agent-ready knowledge graph — deterministically, with no LLM and no API key. It's a reference implementation of both sides of Google's Open Knowledge Format: producing OKF bundles and consuming them.

Install

uvx okfgen generate .        # zero-install
pip install okfgen           # or: pip install "okfgen[all]"

Highlights

Producers — make a bundle from your data

• git, local directory, database schema (schema: JSON/SQL), BigQuery (bq:), Firebase/Firestore (firebase:), web docs, and live open-data portals via CKAN (ckan:) and Socrata (socrata:) — no auth, works against data.gov, NYC, Toronto, and thousands of gov portals.
• The core runs on the Python standard library alone — zero required third-party deps. Cloud SDKs are optional extras, loaded lazily.

Enrichment agent (pass 2)

• Infers foreign-key join paths between tables (customer_id → customers) and wires backlinks so the graph is navigable both ways.

Consumers — read any bundle back

• visualize → a self-contained interactive graph (one HTML file, no backend, data never leaves the page)
• search → a TF-IDF full-text index
• ask → a reasoning agent that answers with citations and shows the concepts it traversed
• validate → OKF v0.1 conformance check

Use it inside your AI agent (MCP)

• Ships an MCP server (okfgen-mcp) exposing generate/search/ask/validate/visualize as tools for Claude Desktop, Claude Code, and Cursor.

Deterministic by default — reproducible facts, not hallucinations; runs offline. An optional --llm flag adds Claude-written polish where you want it.

Try it

uvx okfgen generate . -o my-okf
uvx okfgen visualize my-okf -o my-okf/graph.html

Links

• 🎨 Live demo gallery: https://bushans.github.io/okfgen/
• 📦 PyPI: https://pypi.org/project/okfgen/
• 📖 Getting started: TESTING.md

Notes

• The OKF spec is a v0.1 draft; okfgen tracks it and includes an automated spec-drift watcher.
• Feedback very welcome — especially on the producer adapters and where the deterministic extraction falls short.

Full Changelog: https://github.com/bushans/okfgen/commits/v0.1.0