Photo
CamScanner-style filter strip (Original, Omnifix, Enhance, No-shadow, Eco, B&W, …), AI deblur / enhance text / remove background, default 1280px JPEG q65 preset.
The product
CaptureX — the AI capture layer
A modern reincarnation of Catch Notes (3banana, 2010–2013), rebuilt for multimodal AI. Text, images, audio, video, forwarded chats — captured, understood, organized, and mirrored as plain markdown into a git repository you own.
Plaza Orchestration v1
Every capture earns the same 4-section shape.
Inside CaptureX BO, every persisted capture body is composed of exactly four top-level H2 sections, in canonical order.
## Audit Trail— the steps and models that ran## Capture File— source, attachments, geo, lineage tags## Summary— one paragraph, written by the crafter## Note— the reusable content, structured with H3+
The schema is enforced at the BO layer (composePlazaCaptureNoteBody), defended at the mobile chat-render layer, and asserted by the auditor before persistence.
multi-realm
Realms keep your worlds apart.
A realm is your personal partition: work, research, family, side-project, client A, client B. Each realm has its own Plaza skill bundle, its own default Inbox and Memory space, its own mirror folder, and its own data driver.
- Synthetic Inbox always short-circuits orchestration — raw capture, no router
- Synthetic Memory runs a single memory-organizer skill
- Custom spaces opt-in to multi-space fan-out (up to 4 results)
- Switch realms anywhere; AppBar auto-loads the right skill plaza
R
Realm · work
R
Realm · research
R
Realm · family
All three coexist; one click in the AppBar switches the active realm, refreshes the skill plaza, and re-scopes search.
local-first
One folder, your entire knowledge base.
Pick a realm to use the local_git data driver and CaptureX will mirror that realm into a single absolute folder on disk — as plain markdown, with a deterministic layout.
- Full coverage — captures, spaces, skills, books, tags, compilations
- No auto-commit, no auto-push — operator-driven SCM panel
- Auto pull / fetch keeps the working tree current
- Batched, resumable import scales to 10k+ captures
- Bring your own remote (GitHub, Gitea, GitLab, file:// share)
The git mirror is the canonical store in v1 — read it with any text editor, version it like code, and back it up wherever you back up your code.
books & wiki
From scattered captures to a governed wiki.
Books are first-class long-form documents inside every realm. Feed captures into nested chapters, let the RealmAgent compose and refine content, and mirror the full tree back to git — reversible, diffable, agent-maintainable.
- Wiki bookType for curated reference knowledge bases — lecture digests, paper notes, teaching corpora
- Nested chapters with domain prompts, feed / refine / rewrite modes, and source-capture lineage
- Auto TOC, heading numbering templates, and special pages (cover, abstract, preface, index)
- CaptureW Books panel — outline tree, full-screen editor, chapter search, one-click open
- Full git mirror under books/ with manifest, chapters, and format spec for LLM agents
1
Capture lectures & papers
2
Create a wiki Book
3
Feed captures into chapters
4
Mirror to git as OKF concepts
OKF v0.1
Google Open Knowledge Format, native in your git folder.
Every local_git realm folder is a self-contained OKF v0.1 bundle. Captures, spaces, tasks, memories, sessions, links, and books are typed concept documents with YAML frontmatter — so external tools and LLM agents can read, validate, and evolve your knowledge without proprietary lock-in.
- Root index.md declares okf_version 0.1; _okf/CONFORMANCE.md documents every section
- Every non-reserved .md file carries non-empty type (LarkCatch Capture, Book, Chapter, Space, …)
- Books subtree adds index.md, _OKF_EVOLUTION.md, and per-book indexes with okf_profile metadata
- Legacy imports stay tolerant; re-encode upgrades files to OKF without changing stable ids
- JSON sidecars (skill-plaza.json, compilations.json) remain machine-readable companions
architecture
Built for operators who own their stack.
CaptureX is not a generic note app with a bolt-on sync layer. The backend (CaptureX BO) is a NestJS service you run locally or on your own infra — mobile and desktop clients talk to it over HTTP only.
- Identity via your Service Profile — no shared SaaS tenant
- Storage (v1) = per-realm git repository you choose
- Ingestion = mobile FAB, desktop, share-sheet, capture API
- Event hooks for messenger bots and custom connectors (roadmap)
- App secrets live in a Service Profile JSON, never in env vars
Pair it with your own LLM keys and the deployment is one operator running python ops.py restart.
①
Capture clients
②
CaptureX BO (NestJS)
③
Mobile + Desktop apps (Flutter)
④
Per-realm git repository
Multimodal capture
Photo-first, but every modality is first class.
Receipts, whiteboards, voice memos, screenshots, forwarded chats, PDFs — all funnel through the same orchestrator with the right vendor and the right prompt.
Audio
Voice memos transcribed and summarized; cloud TTS through BO HTTP only — default Volcengine / Doubao, OpenAI-compat fallback. Read-aloud on every capture.
Documents
Drag, paste, or share PDFs and Office docs into the AgentSheet composer. BO extracts text + structure; the auditor flags low-quality extraction.
Forwarded chats
Forward a chat thread from any messenger and it lands as a capture with the original quotes preserved, the senders tagged, and the action items extracted.
Geo & lineage
Photo EXIF GPS + text auto-parse + map-picker (Google / Apple / Amap / inline OSM). Every capture remembers where it came from.
Vendor-neutral
Vision LLM through a vendor-slot (Gemini default), TTS through Volcengine/Doubao with OpenAI-compat fallback, and keys via Service Profile or external secrets file — never client-to-vendor.
scale & durability
From 10 captures to 10,000+ on the same engine.
We design for the long arc. The GET /v1/captures endpoint supports cursor pagination with stub / full include modes; the local-git driver imports in batches with a resumable checkpoint.
- Cursor pagination (1..200) · keyset index per realm/space
- Stub mode omits heavy body fields; detail endpoint hydrates on demand
- Async sync session: start, poll progress, cancel via DELETE
- Re-runs preserve attachments and lineage — no silent data loss
- Revision log appends, never overwrites