Currents

A calm, ad-free visual inspiration platform built on the AT Protocol. Users save images into Collections.

Services

Directory	Language	Description
appview/	Go	AT Protocol AppView — HTTP server with OAuth, TAP listener, backed by PostgreSQL + pgvector
inference/	Python	FastAPI inference server — embeds images and text into a shared vector space (SigLIP2)

The appview subscribes to a TAP WebSocket at startup to index is.currents.* records from the AT Protocol firehose into local collection and save tables. Saves are always written to the user's PDS first (via the browser extension or the appview's POST /resave endpoint for resaves), then indexed asynchronously by the TAP listener — the appview DB is never written to directly by HTTP handlers for user content.

Quick start

SESSION_SECRET=<random-string> docker compose up --build

• AppView: http://localhost:8080
• Inference: run separately (see inference/)

Full local dev stack

./dev.sh starts everything — the dev compose stack (db/tap/appview/clustering), the inference server, the frontend, and a cloudflared tunnel — each in its own tmux window (docker, inference, frontend, tunnel).

./dev.sh          # build + start everything, attach to tmux
./dev.sh stop     # stop inference/frontend/tunnel (docker stack keeps running)
./dev.sh stop --down  # also tear down the docker stack
./dev.sh attach   # reattach after detaching (Ctrl-b d)
./dev.sh restart frontend   # bounce a single window

Dev TAP stack

For local TAP debugging, use the dev compose file directly. It tracks only the handles in TAP_DEV_HANDLES, seeds TAP via /repos/add, and starts the full local stack, including clustering.

TAP_DEV_HANDLES="matteomarjanovic.com,zeroassembly.site" \
docker compose -f docker-compose.dev.yml up -d --build

Add more tracked repos by appending more comma-separated handles to TAP_DEV_HANDLES.

To wipe the dev stack completely:

docker compose -f docker-compose.dev.yml down -v

Inference server

cd inference
pip install -r requirements.txt
uvicorn main:app --reload

Endpoints: POST /embed/image, POST /embed/text, POST /analyze/image, GET /health

/analyze/image returns an embedding, dominant-color palette, and image dimensions in one call — used by the appview when indexing new saves.

Moderation

Currents ships with a first-class atproto labeler. New saves are auto-classified for NSFW / violence / AI-generated content inside the same SigLIP2 pass that produces the embedding, suspect content lands in the moderation review queue at /moderation/queue, and reviewer decisions produce signed labels exposed via com.atproto.label.subscribeLabels / queryLabels so Bluesky clients honor them too.

• Architecture deep-dive: MODERATION.md
• Deployment walkthrough (keypair, DNS, publishing the service record): MODERATION_DEPLOYMENT.md

More

See appview/README.md for full AppView configuration and OAuth docs.