folio-kit 0.1.8

Portable, AI-native data sheets.

folio-kit

Portable, AI-native data sheets. The Python distribution ships three console scripts and a Python SDK for reading and writing typed data sheets that humans and AI agents can both edit.

my-sheet/
├── contract.yaml         # required — ODCS subset
├── records.jsonl         # required — one JSON object per line
├── derivations/          # optional — derivation files (ai / python / sql / cross_sheet / import / http)
├── scripts/              # optional — reusable Python scripts
├── skills/               # optional — packaged operating procedures (markdown + YAML frontmatter)
├── provenance.jsonl      # append-only audit log
└── README.md             # optional, with typed frontmatter

What you get

Surface	Command	Purpose
CLI	folio	init, validate, query, list, count, upsert, delete, materialize, status, provenance, serve, script, skill, export
MCP server	folio-mcp	FastMCP server exposing the SDK over stdio or HTTP — plus one MCP prompt per skill — drop into Claude Desktop / Cursor / any MCP client
Viewer backend	folio-viewer	Local-only FastAPI server that powers the Folio Viewer desktop app and web UI
Python SDK	from folio import open_sheet	Programmatic access to the same operations as the CLI

Install

pipx install folio-kit
# or
uv tool install folio-kit

The distribution name is folio-kit; the on-disk command is folio.

Quickstart

folio validate ./my-sheet
folio materialize ./my-sheet --actor agent:demo
folio query ./my-sheet "SELECT id, status FROM records LIMIT 5"

Programmatic:

from folio import open_sheet

sheet = open_sheet("./my-sheet", actor="agent:demo")
sheet.upsert({"id": "cust_001", "company_name": "Acme", "country": "Japan"})
result = sheet.materialize()
print(result)   # {"materialized": 1, "skipped": 0, "failures": [], "total_cost": 0.0}

Derivations

A derivations/<field>.yaml declares how a derived column is filled. Folio ships six kinds:

• ai — call an LLM (Anthropic by default; injectable AIClient)
• python — run a reusable script from scripts/
• sql — DuckDB query against the sheet's records view
• cross_sheet — pull a value from a sibling Folio sheet (1:1 by PK)
• import — copy from a CSV / JSON file
• http — fetch from a URL

Every value comes with a provenance entry (actor, timestamp, input hash, model + cost for ai). Re-runs are content-hashed and cached.

Requirements

• Python 3.13+
• Optional: ANTHROPIC_API_KEY in the environment if you use kind: ai derivations against the default adapter. Bring-your-own AIClient for any other provider.

Links

• Source & full documentation: https://github.com/nyuta01/folio
• Specification: SPECIFICATION.md — the complete contract, formats, and semantics
• Examples: five working sheets under examples/ (customers + its customer-revenue sidecar, research-notes, research-memory, onboarding) — each ships at least one skills/*.md operating procedure
• Agent skills: the folio-agent-skills npm package ships ready-to-install SKILL.md files for Claude / Cursor / any skills.sh-compatible agent runtime.
• Issues & roadmap: https://github.com/nyuta01/folio/issues

License

MIT.