Plaza Orchestration
Every capture runs through a 4-step pipeline — router, crafter, auditor, tagger — and is persisted with the canonical ## Audit Trail / ## Capture File / ## Summary / ## Note schema.
Building the AI capture layer · 2026
Capture anything.
Capture anything.
Keep it yours. In git.
CaptureX turns every message, photo, document and shared snippet into a structured, AI-organized capture — stored in your own git repositories, governed by your own multi-realm workspace, and indexed for instant retrieval.
-
Receipt — Lao Sichuan dinner 20260524
-
Whiteboard photo — Q3 roadmap 20260523
-
Chat thread — onboarding feedback 20260522
-
Voice memo — investor follow-up 20260521
Structured output · ## Note
Q3 roadmap whiteboard
Three swim-lanes: ingest, recall, govern. Owner sketched a one-month spike for shared-realm Git mirror plus a longer arc for structured export.
## Audit Trail · router → tagger → crafter → auditor · 4 steps · 12.4s
multimodal · vision
4 actions extracted
10k+
captures per realm, scaled & paginated
4-step
Plaza Orchestration · router → outcome
100%
data lives in git repositories you control
3 surfaces
Mobile, Desktop, BO Admin · same source
Why CaptureX
An AI-native capture layer built for deep work
Stop juggling Notion, Apple Notes, Bear and a dozen chat threads. CaptureX unifies the capture surface and pushes everything through a four-section schema that makes future-you searchable.
Multi-realm workspace
Separate work, research, family and side-projects into realms — each with its own Plaza skill bundle, default Inbox, Memory space, and Git-mirrorable disk path.
Capture from anywhere
Mobile FAB, desktop hotkey, share-sheet, paste, and messenger forwarding — every channel funnels into the same ingest endpoint. Storage stays in git repositories you own.
Photo-first capture
Camera-mode FAB, CamScanner-style filter strip, AI denoise / deblur / enhance text, and a default 1280px JPEG-q65 compression preset — receipts, whiteboards, books, all clean in seconds.
Local-first & portable
Per-realm local_git driver mirrors your full workspace — captures, skills, spaces, audit logs — as markdown in one folder. Bring your own remote, your own backup, or stay fully offline.
Vendor-neutral LLM
Multimodal pipeline runs against pluggable vendors (Gemini default, with OpenAI-compatible fallback) — and TTS routes through Volcengine / Doubao via a single BO endpoint, never client-to-vendor.
How it works
From a glance to a governed note in four moves
Capture in the app, on desktop, or via share-sheet. CaptureX BO handles the rest: parsing, routing, summarizing, persisting.
01 · Capture
Catch it where it lives
Mobile FAB, desktop hotkey, share-sheet, paste, or messenger forward — every channel funnels into the same ingest endpoint.
02 · Understand
Multimodal LLM pipeline
Router decides the space, crafter writes the note, auditor checks integrity, tagger labels lineage — vision-passthrough on photos and PDFs.
03 · Persist
Your own git, on your own disk
The canonical body is written as plain markdown into a git repository you choose — durable, diffable, exportable. Push to GitHub, Gitea, GitLab, or keep it strictly local.
04 · Recall
Find anything, anywhere
Full-text + tag + space + lineage search. Open a capture months later and the audit trail tells you exactly what the model thought and saw.
data sovereignty
Your data, in a git repo you own.
CaptureX does not host your notes on our servers. We build the ingestion, parsing and orchestration layer; you own the storage. Every capture, attachment, revision and tag is mirrored into a per-realm git repository on a folder you choose.
- Service Profile JSON is the only source of runtime config
- Per-realm git mirror writes captures as plain markdown
- Secrets stay on your machine — never in a shared cloud vault
- No telemetry, no cross-tenant aggregation, no shadow database
- Plain markdown you can diff, grep, and version like code
deep work
Plaza Orchestration — one schema for every capture.
We dropped the freeform "AI summary" approach. Every persisted body has exactly four H2 sections in canonical order, every tag is governed, and every revision is appended — never overwritten.
## Audit Trail— exactly which steps and models ran## Capture File— source, attachments, geo, lineage## Summary— one paragraph the model would defend## Note— the actual reusable content
1
Router
2
Outcome Crafter
3
Auditor
4
Lineage Tagger