feat: caption-ninja recipe (web pages, private relay, Caption Local STT)

Self-hosted live captioning from steveseguin/captionninja and
steveseguin/caption-local. Upstream publishes no images, so images/build.sh
builds them from pinned commits.

.env.sample

@@ -0,0 +1,39 @@
+TYPE=caption-ninja
+# First deploy downloads the Whisper model; give abra time to converge.
+TIMEOUT=900
+
+# caption.ninja pages (capture, editor, overlay) and the private relay at /relay
+DOMAIN=caption-ninja.example.com
+# Caption Local speech-to-text service (needs its own hostname: it uses absolute paths)
+STT_DOMAIN=stt-caption-ninja.example.com
+
+LETS_ENCRYPT_ENV=production
+
+# Relay room tokens: 32-128 URL-safe chars, all four must differ.
+# "source" = capture -> editor, "output" = editor -> overlay/OBS.
+SECRET_SRC_READ_VERSION=v1 # length=48
+SECRET_SRC_WRITE_VERSION=v1 # length=48
+SECRET_OUT_READ_VERSION=v1 # length=48
+SECRET_OUT_WRITE_VERSION=v1 # length=48
+# Caption Local shared service token (>= 24 chars, no whitespace)
+SECRET_STT_API_KEY_VERSION=v1 # length=48
+
+## Speech-to-text tuning (CPU). See upstream docs/DEPLOYMENT-PROFILES.md.
+## Whisper model: tiny, base, small (default), medium, large-v3. Bigger = more RAM + CPU.
+#CAPTION_MODEL=small
+## Decoder beam width 1-5; 1 is fastest
+#CAPTION_BEAM_SIZE=5
+## Concurrent model workers; threads are per worker. Keep workers*threads <= cores.
+#CAPTION_WORKERS=1
+#CAPTION_THREADS=4
+#CAPTION_MAX_STREAMS=12
+## Set to 1 once the model has downloaded into the volume, to stop Hugging Face lookups
+#CAPTION_OFFLINE=0
+## Metadata-only request logs (no audio/text/tokens)
+#CAPTION_LOG_REQUESTS=0
+
+## Extra exact origins (comma separated, e.g. https://steveseguin.github.io)
+#RELAY_EXTRA_ORIGINS=
+#STT_EXTRA_ORIGINS=
+## Max relay sockets per client IP (everyone shares Traefik's IP)
+#RELAY_PER_IP=512

README.md

@@ -0,0 +1,98 @@
+# CAPTION.Ninja + Caption Local
+
+Self-hosted live captioning: [CAPTION.Ninja](https://github.com/steveseguin/captionninja)'s
+capture/editor/overlay pages and private caption relay, plus
+[Caption Local](https://github.com/steveseguin/caption-local) Whisper speech-to-text.
+Nothing is sent to caption.ninja, Google or any cloud STT provider.
+
+<!-- metadata -->
+* **Category**: Apps
+* **Status**: 1, alpha
+* **Image**: [`git.autonomic.zone/recipe-maintainers/caption-ninja-*`](https://git.autonomic.zone/recipe-maintainers/-/packages), 1, built from upstream source (see [Images](#images))
+* **Healthcheck**: Yes
+* **Backups**: No (no user data; the model volume is a cache)
+* **Email**: No
+* **Tests**: 3
+* **SSO**: No
+<!-- endmetadata -->
+
+## Services
+
+| Service | Hostname | What |
+| --- | --- | --- |
+| `app` | `$DOMAIN` | nginx serving the caption.ninja pages |
+| `relay` | `wss://$DOMAIN/relay` | private caption relay (Node, `ws`), rooms `source` and `output` |
+| `stt` | `$STT_DOMAIN` | Caption Local (faster-whisper, CPU), also serves its own capture page |
+
+Captions flow: **capture** (microphone → STT) publishes to room `source` → **editor**
+reviews and forwards to room `output` → **overlay** (OBS browser source / audience).
+
+## Basic usage
+
+1. Set up Docker Swarm, [`abra`][abra] and [`coop-cloud/traefik`][cc-traefik]
+2. Make the images available on the server (see [Images](#images))
+3. `abra app new caption-ninja --secrets`
+4. `abra app config YOURAPPDOMAIN`: set `DOMAIN` and `STT_DOMAIN` (both need DNS).
+   The STT host needs RAM for the model: about 1.5 GB for `small`. Set `CAPTION_THREADS` to at most your core count.
+5. `abra app deploy YOURAPPDOMAIN`. On first start `stt` downloads the Whisper
+   model into the `models` volume, so it can take several minutes to turn healthy.
+
+## Using it
+
+Read the tokens (keep them private; read tokens are for viewers, write tokens publish):
+
+```
+abra app run YOURAPPDOMAIN relay -- sh -c 'for s in src_read src_write out_read out_write; do echo "$s=$(cat /run/secrets/$s)"; done'
+abra app run YOURAPPDOMAIN stt -- cat /run/secrets/stt_api_key
+```
+
+1. Open `https://$STT_DOMAIN/`, enter the `stt_api_key` service token, then pick a microphone.
+2. Under **Send captions to caption.ninja → Use a private relay**, enter:
+   relay `wss://$DOMAIN/relay`, caption website `https://$DOMAIN/`, source room
+   `source` + `src_write` token, editor output room `output`. Enable sharing.
+3. Open the generated **editor** link and enter `src_read` and `out_write`.
+4. In the editor, create the view-only OBS/overlay link with `out_read`.
+
+Alternatively use `https://$DOMAIN/capture-local.html` with service address
+`https://$STT_DOMAIN`.
+
+**Note:** the other caption.ninja pages (`index.html`, `translate.html`, …) still
+default to the public relay `wss://api.caption.ninja`. Only `capture-local.html`,
+`editor.html` and `overlay.html` support the private relay.
+
+## Images
+
+Upstream publishes no Docker images, so `images/build.sh` builds them from pinned
+upstream commits. It needs Docker, git and no registry credentials:
+
+```
+scp -r images SERVER:caption-ninja-build/
+ssh SERVER 'bash caption-ninja-build/images/build.sh'
+```
+
+The images are tagged with their registry names, so a single-node swarm can deploy
+straight from the local tags. To publish them to the registry from a machine
+that has `docker login git.autonomic.zone`:
+
+```
+for img in $(ssh SERVER "docker images --format '{{.Repository}}:{{.Tag}}' | grep caption-ninja-"); do
+  ssh SERVER docker save "$img" | docker load && docker push "$img"
+done
+```
+
+To upgrade, bump the commits in `images/build.sh` and the tags in `compose.yml`.
+
+## Security notes
+
+* Caption Local only accepts `Host: localhost` (upstream expects an SSH tunnel), so
+  Traefik rewrites the Host header for `$STT_DOMAIN`; `stt` is still only reachable
+  through Traefik, and every API call needs the token.
+* The STT token is a single shared token, not per-user auth. Anyone with it can use
+  your CPU for transcription. Rotate it with `abra app secret` + redeploy.
+* Relay origins are `https://$DOMAIN` and `https://$STT_DOMAIN`; add more exact
+  origins with `RELAY_EXTRA_ORIGINS` (e.g. `https://steveseguin.github.io`).
+* To revoke relay tokens, generate new secret versions and redeploy (restart
+  disconnects all clients and clears the 2-minute replay buffer).
+
+[abra]: https://git.coopcloud.tech/coop-cloud/abra
+[cc-traefik]: https://git.coopcloud.tech/coop-cloud/traefik

abra.sh

@@ -0,0 +1,3 @@
+export NGINX_CONF_VERSION=v1
+export RELAY_ENTRYPOINT_VERSION=v1
+export STT_ENTRYPOINT_VERSION=v2

compose.yml

@@ -0,0 +1,153 @@
+---
+version: "3.8"
+
+services:
+  app:
+    image: git.autonomic.zone/recipe-maintainers/caption-ninja-web:2026.09.07-dfcee8f
+    networks:
+      - proxy
+    configs:
+      - source: nginx_conf
+        target: /etc/nginx/conf.d/default.conf
+    healthcheck:
+      test: ["CMD", "wget", "-q", "-O", "/dev/null", "http://127.0.0.1/"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+    deploy:
+      update_config:
+        failure_action: rollback
+        order: start-first
+      labels:
+        - "traefik.enable=true"
+        - "traefik.http.routers.${STACK_NAME}.rule=Host(`${DOMAIN}`)"
+        - "traefik.http.routers.${STACK_NAME}.entrypoints=web-secure"
+        - "traefik.http.routers.${STACK_NAME}.tls.certresolver=${LETS_ENCRYPT_ENV}"
+        - "traefik.http.services.${STACK_NAME}.loadbalancer.server.port=80"
+        - "coop-cloud.${STACK_NAME}.version=0.1.0+v2026.09.07"
+
+  relay:
+    image: git.autonomic.zone/recipe-maintainers/caption-ninja-relay:0.2.0-dfcee8f
+    networks:
+      - proxy
+    environment:
+      - DOMAIN
+      - STT_DOMAIN
+      - RELAY_EXTRA_ORIGINS
+      - RELAY_PER_IP
+    secrets:
+      - src_read
+      - src_write
+      - out_read
+      - out_write
+    configs:
+      - source: relay_entrypoint
+        target: /relay-entrypoint.sh
+        mode: 0555
+    entrypoint: /relay-entrypoint.sh
+    healthcheck:
+      test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8787/health').then(r=>process.exit(r.ok?0:1),()=>process.exit(1))"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+    deploy:
+      update_config:
+        failure_action: rollback
+        order: start-first
+      labels:
+        - "traefik.enable=true"
+        - "traefik.http.routers.${STACK_NAME}_relay.rule=Host(`${DOMAIN}`) && Path(`/relay`)"
+        - "traefik.http.routers.${STACK_NAME}_relay.entrypoints=web-secure"
+        - "traefik.http.routers.${STACK_NAME}_relay.tls.certresolver=${LETS_ENCRYPT_ENV}"
+        - "traefik.http.routers.${STACK_NAME}_relay.service=${STACK_NAME}_relay"
+        - "traefik.http.routers.${STACK_NAME}_relay.middlewares=${STACK_NAME}_relay_strip"
+        - "traefik.http.middlewares.${STACK_NAME}_relay_strip.stripprefix.prefixes=/relay"
+        - "traefik.http.services.${STACK_NAME}_relay.loadbalancer.server.port=8787"
+
+  stt:
+    image: git.autonomic.zone/recipe-maintainers/caption-ninja-stt:1.1.0-7d759fe
+    networks:
+      - proxy
+    environment:
+      - DOMAIN
+      - STT_DOMAIN
+      - STT_EXTRA_ORIGINS
+      - CAPTION_MODEL=${CAPTION_MODEL:-small}
+      - CAPTION_BEAM_SIZE=${CAPTION_BEAM_SIZE:-5}
+      - CAPTION_WORKERS=${CAPTION_WORKERS:-1}
+      - CAPTION_THREADS=${CAPTION_THREADS:-4}
+      - CAPTION_MAX_STREAMS=${CAPTION_MAX_STREAMS:-12}
+      - CAPTION_COMPUTE_TYPE=${CAPTION_COMPUTE_TYPE:-auto}
+      - CAPTION_DEVICE=cpu
+      - CAPTION_OFFLINE=${CAPTION_OFFLINE:-0}
+      - CAPTION_LOG_REQUESTS=${CAPTION_LOG_REQUESTS:-0}
+    secrets:
+      - stt_api_key
+    configs:
+      - source: stt_entrypoint
+        target: /stt-entrypoint.sh
+        mode: 0555
+    entrypoint: /stt-entrypoint.sh
+    volumes:
+      - models:/models
+    healthcheck:
+      # The image's own healthcheck reads CAPTION_API_KEY from the environment,
+      # which only the entrypoint process has; feed it the secret here.
+      test: ["CMD", "sh", "-c", "CAPTION_API_KEY=$$(cat /run/secrets/stt_api_key) python healthcheck.py"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      # First start downloads the Whisper model into the volume.
+      start_period: 600s
+    stop_grace_period: 100s
+    deploy:
+      update_config:
+        failure_action: rollback
+        order: stop-first
+      labels:
+        - "traefik.enable=true"
+        - "traefik.http.routers.${STACK_NAME}_stt.rule=Host(`${STT_DOMAIN}`)"
+        - "traefik.http.routers.${STACK_NAME}_stt.entrypoints=web-secure"
+        - "traefik.http.routers.${STACK_NAME}_stt.tls.certresolver=${LETS_ENCRYPT_ENV}"
+        - "traefik.http.routers.${STACK_NAME}_stt.service=${STACK_NAME}_stt"
+        # Caption Local only accepts Host: localhost (upstream expects an SSH tunnel)
+        - "traefik.http.routers.${STACK_NAME}_stt.middlewares=${STACK_NAME}_stt_host"
+        - "traefik.http.middlewares.${STACK_NAME}_stt_host.headers.customrequestheaders.Host=localhost"
+        - "traefik.http.services.${STACK_NAME}_stt.loadbalancer.server.port=8765"
+
+secrets:
+  src_read:
+    name: ${STACK_NAME}_src_read_${SECRET_SRC_READ_VERSION}
+    external: true
+  src_write:
+    name: ${STACK_NAME}_src_write_${SECRET_SRC_WRITE_VERSION}
+    external: true
+  out_read:
+    name: ${STACK_NAME}_out_read_${SECRET_OUT_READ_VERSION}
+    external: true
+  out_write:
+    name: ${STACK_NAME}_out_write_${SECRET_OUT_WRITE_VERSION}
+    external: true
+  stt_api_key:
+    name: ${STACK_NAME}_stt_api_key_${SECRET_STT_API_KEY_VERSION}
+    external: true
+
+configs:
+  nginx_conf:
+    name: ${STACK_NAME}_nginx_conf_${NGINX_CONF_VERSION}
+    file: nginx.conf
+  relay_entrypoint:
+    name: ${STACK_NAME}_relay_entrypoint_${RELAY_ENTRYPOINT_VERSION}
+    file: relay-entrypoint.sh
+  stt_entrypoint:
+    name: ${STACK_NAME}_stt_entrypoint_${STT_ENTRYPOINT_VERSION}
+    file: stt-entrypoint.sh
+
+networks:
+  proxy:
+    external: true
+
+volumes:
+  models:

images/build.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Build the three caption-ninja images from pinned upstream commits.
+#
+# Upstream publishes no images, so we build them ourselves. Run this on a
+# Docker host (it needs no registry credentials); the images are tagged with
+# their final registry names so a single-node swarm can deploy them directly.
+# See README.md "Images" for pushing them to the registry afterwards.
+#
+#   usage: images/build.sh [workdir]
+set -euo pipefail
+
+REGISTRY=git.autonomic.zone/recipe-maintainers
+CAPTIONNINJA_COMMIT=dfcee8f551a2b175740ab9f73e3787e1458cce2d
+CAPTION_LOCAL_COMMIT=7d759fe1eae100ccdb5de0eea82438c941b26582
+RELAY_VERSION=0.2.0       # captionninja relay/package.json
+CAPTION_LOCAL_VERSION=1.1.0  # caption-local VERSION
+WEB_VERSION=2026.09.07    # captionninja has no releases: date of CAPTIONNINJA_COMMIT
+
+here="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+work="${1:-$(mktemp -d)}"
+
+checkout() {  # url commit dir
+    rm -rf "$3"
+    git init -q "$3"
+    git -C "$3" fetch -q --depth 1 "$1" "$2"
+    git -C "$3" checkout -q FETCH_HEAD
+}
+
+checkout https://github.com/steveseguin/captionninja "$CAPTIONNINJA_COMMIT" "$work/captionninja"
+checkout https://github.com/steveseguin/caption-local "$CAPTION_LOCAL_COMMIT" "$work/caption-local"
+
+cn="${CAPTIONNINJA_COMMIT:0:7}"
+cl="${CAPTION_LOCAL_COMMIT:0:7}"
+
+docker build -t "$REGISTRY/caption-ninja-web:$WEB_VERSION-$cn" \
+    -f "$here/web/Dockerfile" "$work/captionninja"
+docker build -t "$REGISTRY/caption-ninja-relay:$RELAY_VERSION-$cn" \
+    "$work/captionninja/relay"
+docker build -t "$REGISTRY/caption-ninja-stt:$CAPTION_LOCAL_VERSION-$cl" \
+    "$work/caption-local"
+
+echo
+echo "Built:"
+echo "  $REGISTRY/caption-ninja-web:$WEB_VERSION-$cn"
+echo "  $REGISTRY/caption-ninja-relay:$RELAY_VERSION-$cn"
+echo "  $REGISTRY/caption-ninja-stt:$CAPTION_LOCAL_VERSION-$cl"

images/web/Dockerfile

@@ -0,0 +1,7 @@
+# Static caption.ninja pages. Build context: a captionninja checkout.
+FROM alpine:3.22 AS src
+COPY . /src
+RUN cd /src && rm -rf .git .github relay scripts *.md test-security-warning.html
+
+FROM nginx:1.31.6-alpine
+COPY --from=src /src /usr/share/nginx/html

nginx.conf

@@ -0,0 +1,15 @@
+server {
+    listen       80;
+    listen  [::]:80;
+    server_name  _;
+    root   /usr/share/nginx/html;
+    index  index.html;
+
+    # caption.ninja links omit the extension (e.g. /overlay?room=...).
+    location / {
+        try_files $uri $uri.html $uri/ =404;
+    }
+
+    # Room tokens can arrive in query strings on some pages; keep them out of logs.
+    access_log off;
+}

relay-entrypoint.sh

@@ -0,0 +1,28 @@
+#!/bin/sh
+set -e
+
+# The relay reads one private JSON file holding allowed origins and per-room
+# tokens. Build it from the Docker secrets at startup so tokens never live in
+# the recipe or the env file. Writes to /tmp, readable only by the node user.
+umask 077
+node - > /tmp/rooms.json <<'JS'
+const fs = require('node:fs');
+const secret = name => fs.readFileSync('/run/secrets/' + name, 'utf8').trim();
+const origins = [process.env.DOMAIN, process.env.STT_DOMAIN]
+  .filter(Boolean).map(host => 'https://' + host)
+  .concat((process.env.RELAY_EXTRA_ORIGINS || '').split(',').map(s => s.trim()).filter(Boolean));
+const config = {
+  origins: [...new Set(origins)],
+  rooms: {
+    source: {read: secret('src_read'), write: secret('src_write')},
+    output: {read: secret('out_read'), write: secret('out_write')},
+  },
+  // Every client arrives via Traefik, so they all share one source IP; the
+  // upstream default of 128 sockets per IP would cap the whole audience.
+  limits: {perIp: Number(process.env.RELAY_PER_IP || 512)},
+};
+process.stdout.write(JSON.stringify(config));
+JS
+
+export CAPTION_RELAY_CONFIG=/tmp/rooms.json
+exec node server.cjs

stt-entrypoint.sh

@@ -0,0 +1,16 @@
+#!/bin/sh
+set -e
+
+# Caption Local takes its shared service token from CAPTION_API_KEY only.
+if [ -f /run/secrets/stt_api_key ]; then
+    export CAPTION_API_KEY="$(cat /run/secrets/stt_api_key)"
+fi
+
+# Let the caption.ninja pages on $DOMAIN call this service, plus any extras.
+# Its own page on $STT_DOMAIN must be listed too: Traefik rewrites Host to
+# localhost, so the app no longer sees that page as same-origin.
+origins="https://${DOMAIN},https://${STT_DOMAIN}"
+[ -n "$STT_EXTRA_ORIGINS" ] && origins="${origins},${STT_EXTRA_ORIGINS}"
+export CAPTION_ALLOWED_ORIGINS="$origins"
+
+exec python server.py --host 0.0.0.0 "$@"