Overview

ScaiScribe is the ScaiLabs document service. You hand it a JSON document specification — headings, paragraphs, tables, charts, slides, sheets — and get back DOCX, PPTX, XLSX, or PDF artefacts. You can also go the other way: hand it a DOCX and get back a specification you can edit, then re-render.

The same backend powers all three workflows: author from scratch, template-fill with placeholders, and ingest-and-edit existing documents.

What it's good for#

Programmatic document generation — invoice runs, monthly reports, personalised PDFs. A spec is structured data, so you can author it from any source (database row, LLM call, form submission).
AI agents — the MCP surface is the primary interface. An agent calls scribe_create_document, scribe_add_element, scribe_finalize and gets a download URL back.
Template instantiation — register a .docx with {{placeholder}} markers, fill it with structured data per request.
Ingest + edit — round-trip an existing DOCX through the spec format. Useful for "extract content from this contract, modify clause 3, re-emit" pipelines.

What you talk to#

Three peer surfaces, all backed by the same service layer:

Surface	Best for
MCP (46 tools)	AI agents. `scribe_*` tools cover the full lifecycle.
REST (`/v1/*`)	Application code in any language. Bearer auth, JSON in/out.
JSON spec format	Power-user escape hatch — `POST /v1/documents/from_spec` with a hand-written spec.

The published spec at schema/scaiscribe.schema.json is the canonical contract; both Python and TypeScript SDKs ship types generated from it.

What it can't do today (alpha caveats)#

We're transparent about what's deferred to v1.x so you don't trip over it:

DOCX/XLSX native charts for kinds outside bar/column/line/pie — those four ship as hand-emitted OOXML chart parts (editable in Word/Excel). Other kinds (area, scatter, sankey, treemap, gauge, etc.) render as ECharts SVG images — visually correct but not editable as native charts. PPTX native is broader (covers pptxgenjs's full surface).
OCR ingestion — text-only ingest for now.
ScaiSign / ScaiCMS bridge — coming with their respective integrations.

Every deferral lands in the response as a structured warning, never silently dropped.

How auth works#

ScaiScribe authenticates users and services through ScaiKey. For SDK usage:

python
auth = ScaiKeyAuth(client_id="...", client_secret="...")
client = Client("https://scaiscribe.scailabs.ai", auth=auth)

The SDK handles token fetch + refresh transparently. See Quickstart for the full flow.

Where to next#

Just want to render a document? Go to Quickstart — sub-five-minute path from zero to a downloaded DOCX.
Building an AI agent? The MCP reference lists every scribe_* tool.
Integrating from code? The SDK reference covers Python, .NET, and TypeScript.

Status#

ScaiScribe is in alpha at https://scaiscribe.scailabs.ai. Pin SDK versions and expect minor breaking changes through v1.0 GA. Production-shaped surface is stable; deferrals are documented per page.