00.00 suchi · an index that keeps itself

The document archive that files itself.

Years of documents, saved everywhere, findable nowhere. suchi ends the pile: everything that comes in files itself under a number that never moves, and every word of it is searchable. One small binary, with a filing system and an AI agent surface built in.

email it, scan it, upload it. it comes out named, filed, and searchable

Get started in five minutes See what it does

Private by design Share links you can revoke No lock-in Runs on a Pi

21Banking38 docs

22Tax+1 filed

23Insurance17 docs

24Investments52 docs

now filing · original kept as-is IMG_2047.pdf → 20-29 Money / 22 Tax · “ITR-1, AY 2026-27”

00–09 See it

Six seconds, start to filed.

A scan arrives, the index row draws itself, the category stamps on, and the row slides into its numbered place between its neighbours.

Try the live demo → See everything it does → Read the docs →

suchi · dropped in, filed, found

suchi dashboard: metric cards, recently added documents, custom views with live counts — **The dashboard.** Live counts, the newest arrivals, and your saved views. Each one is a filter with a number on it.

Documents list in dark mode with thumbnails, multi-select and a bulk action bar — **Working through a pile.** Multi-select then refile, share, or trash the lot in one transaction. Dark mode included.

Document detail with a blurred confidential preview, metadata, similar documents and versions — **One document, everything about it.** Confidential previews stay blurred until you ask. Similar documents surface on their own.

10–19

What it does

One small binary that takes in almost anything, gives it real structure, automates the busywork, and keeps every byte on your hardware.

Everything finds its way in

Watched folders, drag and drop, a sidecar spec for scripts, and email: receive a document from audit@… and it lands filed under 22 Tax. Scans get cleaned up on arrival, and only pages that need OCR pay for OCR.

emailsoffice docsEPUB · DjVu · HEICOutlook .msg + .emle-invoicesblank-page removalauto document splittingQR & barcodesper-page OCR routing

Organized your way

Johnny.Decimal numbering is the opinionated default: browse by number, search jd:2*, nothing is ever unfiled. Prefer classic tags and folders? Flat mode keeps everything you already know.

JD by defaultflat modenested tagscorrespondents with roles8 custom field typesdocument versionstaxonomy merge

It runs itself

Automations fire on trigger, condition, action: tag it, file it, route it. When a human has to sign off, approval chains with timeouts keep the loop transparent, and nothing ever fails silently.

automations engineapproval chainshuman-in-the-loop taskstimeouts & escalationretries with backoffdead-letter visibility

Built for AI agents

Rules handle the obvious. Any OpenAI-compatible model handles the rest, Ollama included, so nothing has to leave the box. Agents get a real seat: an MCP server ships in the binary.

suchi mcpOllama or any endpointrules enginetask-claim loopper-agent tokensevery action audited

Private by construction

A stock install makes zero outbound connections. Sensitive documents can blur their own previews. Sharing is deliberate: expiring links, optional passwords, one switch to disable it all.

zero egresssensitivity labelsblurred previewsshare links: expiry + passwordegress_ack for cloud AIaudit log

20–29

Why an archive, not a folder

You do not need another place to put files. You need to stop losing them. This is what changes when the archive maintains its own index.

Finding a document scroll the folder, guess the filename, give up, ask for a reissue	search every word, or walk ten numbered drawers
Filing a document name it yourself and hope you were consistent in 2019	titled and numbered for you; the original file is never touched
Duplicates IMG_2047.pdf, IMG_2047(1).pdf, IMG_2047 copy.pdf	same bytes stored once, always
What it costs to run a $5 VPS, a Pi, or the corner of a NAS	one process, ≤ 100 MB idle

30–39

Private because of how it’s built

A document archive holds the most sensitive files you own. So the rule is structural: a stock suchi makes zero outbound connections, and every possible egress is opt-in, visible in config, and logged. suchi doctor prints the active list.

What suchi never does

Send telemetry, usage stats, or update pings, not even as an opt-in
Log document content or OCR text, at any log level
Put content in audit events or diagnostics output
Send your documents to a cloud LLM without an explicit egress_ack
Pretend deleted data is gone everywhere. Soft-delete windows and your own backups are named in the docs

Encrypted in transit

TLS at your reverse proxy, or native TLS_CERT_FILE for proxy-less installs. Secure, HttpOnly cookies throughout.

Encrypted volume as the baseline

LUKS or NAS-native encryption covers blobs, database, and thumbnails in one move. The docs recommend it plainly rather than pretending app-layer crypto can substitute for it.

Encrypted backups, for free

restic and borg encrypt client-side by default, so the documented backup story is end-to-end encrypted off-box already.

40–49

For self-hosters and operators

The short version: AGPL-3.0, one compose block, zero external services, runs on a $5 VPS or a Pi, and backups are one directory. An existing archive imports with one command, metadata intact.

docker-compose.yml

# the whole file
services:
  suchi:
    image: ghcr.io/johnnybravo-xyz/suchi:latest
    restart: unless-stopped
    volumes:
      - ./data:/data
    ports:
      - "8000:8000"
    environment:
      PUBLIC_URL: https://docs.example.com
      # everything else has sensible defaults

importing an existing archive

# point it at your current tool's export bundle
$ suchi import \
    --from ./export-bundle \
    --map-jd mapping.yaml

✓ 4,182 documents · metadata intact
✓ folder layout preserved, byte-equal
✓ unmapped docs waiting in 49 Inbox

✓

First boot prints a setup link. No identity provider required; add OIDC later if you run one.

✓

Runs anywhere: compose, bare binary with systemd, NAS templates, arm64, even Windows.