Reelsaw

Self-hosted, AI-powered long-form → short-form video clipper. Paste a YouTube link or upload a video; Reelsaw transcribes it, asks an LLM for the most clippable moments, then renders vertical 9:16 clips with word-highlight captions and a hook overlay. A hosted service at app.reelsaw.com runs the same code (Free, then $12 / $29 / $59 a month, metered in source minutes).

Pure TypeScript. Any AI vendor via the AI SDK. One-step deploy with Docker Compose.

One video through the five queues: ingest → transcribe → refine ×3 → pick moments → render. A worker restarts mid-render and resumes from its last save. Full-quality MP4, or play it on reelsaw.com.

demo-framing.mp4

Auto reframing in one clip: split → face → stack → fit

Features

Pipeline

• Any AI vendor – language model, transcription model and caption-alignment model are each a provider:model string: OpenAI, Anthropic, Google, Groq, Deepgram, AssemblyAI, Vercel AI Gateway, or any OpenAI-compatible server (Ollama, vLLM, LM Studio, local Whisper servers). Fully local with no API keys is supported.
• Research-backed clip selection – map → pool → rank: the transcript is chunked at long pauses with 20% overlap (no "lost in the middle" on long videos), every standalone moment is collected per chunk in parallel, candidates are snapped/deduped, then scored on a rubric (hook / value / flow / shareability, plus relevance with a focus) as the median of k samples, and the top 10 get K sliding pairwise head-to-head passes (PRP-Sliding-K; ANALYZE_PAIRWISE_PASSES, default 3 so the top 3 you see first are placed reliably) with positions swapped to cancel judge bias. Cost: at most Σ(n−1−p) pairs × 2 calls, i.e. 48 calls for the top 10 with 3 passes (18 with 1); pairs met twice are judged once. No target count: in auto mode weak picks are dropped by a gate relative to the run, not a fixed number, because pointwise scores do not transfer between judge models: a clip is dropped below max(ANALYZE_MIN_SCORE_FLOOR = 25, median − ANALYZE_SCORE_MAD_K = 1 × MAD) of the pooled rubric totals, and the best clip is always kept (ANALYZE_SCORE_RULE=absolute restores the old fixed 45). The threshold and its median/MAD are in the project's candidates artifact. Candidates whose scoring failed are never compared with rubric totals; they only fill in when fewer than 3 scored clips survive. Optional "at most N" cap, and "Find clips about …" keeps only on-topic moments. Click a clip's score for the breakdown. ANALYZE_JUDGE_MODEL lets a different model score and compare (less self-preference than judging its own hooks). ANALYZE_CONCURRENCY, ANALYZE_SCORE_SAMPLES, ANALYZE_PAIRWISE tune cost (auto: 1 call at a time, 1 sample, no pairwise for a self-hosted judge; 4 / 3 / on for cloud).
• Clean transcripts – names are fixed from vocabulary hints and a constrained LLM correction, and fillers, repetitions and false starts are cut with jump cuts. Every cut is located by forced alignment and checked by listening to the seam, so karaoke highlights land on the syllable. Each fix shows inline in the transcript with accept/reject, and long pauses can be shortened (details below).
• Smart 9:16 reframing – UltraFace (1.2 MB ONNX, CPU, ~7 ms/frame via onnxruntime-node) samples each clip at 4 fps and picks a layout per segment (clips are split at scene cuts and face-presence changes, so a picture-in-picture intro can be "fit" and the talking head after it "face"; shot cuts are detected on every clip, and a multi-camera edit that alternates a close-up with a wide shot every second or two changes layout at each cut): face (crop follows the speaker with "heavy tripod" smoothing: dead zone, capped pan speed, instant snap at scene cuts), split (two speakers stacked), stack (slides/screen on top, the webcam tile below — the picture-in-picture box of a Zoom, Teams, Meet or OBS recording, not a face-sized crop that reaches into the slides), or fit (whole frame over a blurred background when there is no face, e.g. graphics and b-roll, or when three or more people share the frame). With two people in a wide shot the crop follows whoever is talking (mouth motion correlated with the audio envelope, minimum 2.5 s per shot), falling back to a split when detection isn't confident. Override per project, per clip or per section, including a hand-set crop. Reference for every layout, its triggers, caveats and fixtures: docs/LAYOUTS.md.
• Captions – AI-picked key words per clip coloured and enlarged (editable in the editor), readability limits (≈2 short lines, comma/sentence breaks, a line stays up long enough to read: at least 0.7 s, 0.3 s per word and 20 characters per second), ASS karaoke word highlight (any colour), boxed, or outlined; size, words on screen (1 = one word at a time), uppercase; hook overlay with its own duration. Caption and hook positions are free x/y set by dragging on a live 9:16 preview that uses the renderer's crop math and font metrics, with a platform safe-zone overlay. In split and stack sections that last 2.5 s or more, captions sit on the line between the two panes instead of on the lower speaker (a switch turns it off). Project-wide defaults, per-clip overrides. Burned in with ffmpeg + libass.

Editing

• Workspace and clip editor – source player and clickable transcript on the left, clips streaming in on the right as they render, with score, layout badge, post caption + hashtags copy and download. The editor has a live preview of the source with the clip's framing and captions and a multi-lane timeline (ruler, transcript words, clip range, layout track, layers) with zoom, snapping to words/scene cuts/playhead, nudge buttons, typed timestamps and I/O keys for frame-accurate trims. Split the layout track (⌘/Ctrl+B, the Split button, a double-click or the right-click menu) into sections that are Auto or forced to face/split/stack/fit/centre; drag the picture to re-centre a section; right-click for Set crop… (a box on the source, or crop / crop2 / contentCrop on update_clip); "Use for all clips…" copies one clip's framing to the project (apply_framing). Add, drag and remove transitions; ⌘/Ctrl+Z / ⇧⌘Z undo and redo every edit.
• Deletes can be undone – deleting a project or a clip (card ⋯, the selection bar, the editor, ⌘K) stops its work and hides it. The toast always offers Undo, Delete forever, and See deletions. Everything restorable lives on Deleted (account menu) as one selectable list: each row says when it was deleted and when it is deleted forever, and a compact bar restores the selection or deletes it forever. Storage still counts until you delete it forever (MCP list_deleted plus purge_project / purge_clip / purge_asset / purge_layer_template; reelsaw purge-project, purge-clip, library purge, templates purge). Minutes, hourly creates and first-video-free are never refunded. Re-picking moments keeps every clip you made or edited (text, bounds, framing, style, intro/outro) and only replaces untouched AI picks, which can be restored too. The same through MCP (delete_clip, delete_clips, restore_project, restore_clip, list_deleted) and the CLI (reelsaw deleted, restore, delete-clip, delete-clips, restore-clip, purge, purge-project, purge-clip).
• Learns from your decisions – what you accept, reject, cut, correct and edit is kept in a decision log (with the words or clip text it was about, the models that proposed it and a fingerprint of the source video) that outlives deleted projects: cuts you rejected are shown to the filler detector as examples to keep, and an export turns your decisions into eval cases (reelsaw decisions export, export_decisions, the decisions suite). Settings → Learning turns it off or forgets everything; deleting a project (or "Delete forever") can forget what it learned from that project, applied when it is purged, even automatically. It never leaves your instance. See docs/specs/decision-log.md.
• Command palette and bulk export – ⌘K searches projects and clips with status, length and age, opens a row's actions (open, export, resume, retry failed clips, re-run refine, re-render, download, delete project or clip), takes pasted links (one project each, or opens the existing one) and offers retry-all shortcuts. "Export all" zips every rendered clip with platform-ready metadata: a paste-ready .txt (title, description, tags, caption), a .youtube.json (YouTube Data API v3 videos.insert body, private by default) and a .tiktok.json (Content Posting API post_info, SELF_ONLY) per clip, plus manifest.json, clips.csv (with rendered duration) and a README. Descriptions carry a timestamped link to the exact moment in the parent video.
• SRT / VTT in and out – every rendered clip's ⋯ menu offers its captions as .srt or .vtt (the same lines and times as the video, intro included, even when the style has captions off), and "Export all" adds them next to each clip as clips/NN-title.srt and .vtt; clips rendered before this feature need a re-render first. The project's ⋯ menu ("Transcript files…") downloads the transcript the same way (VTT can carry a timestamp before every word) and imports a subtitle file you already have: right after creating a project (only the video is fetched) or on a done or failed one (it cleans the transcript up and picks clips again). Imported cues become segments and their words are spread over the cue by length; the render re-aligns words to the audio. Also GET /api/clips/:id/captions.srt|vtt, GET /api/projects/:id/transcript.srt|vtt (?words=1 for VTT word times), the MCP tools get_clip_captions, get_project_transcript_file and import_transcript, and reelsaw captions, transcript-file and import-transcript.

Library, layers and templates

• Library – intros and outros, logos, stickers, b-roll, music and sound effects (PNG, JPG, GIF, WebP, SVG, video, audio). Drop many files at once (duplicates are detected), search your items plus a bundled catalog and, with an admin key, Pexels, Pixabay, Freesound, Iconify, Giphy and Klipy. Each clip is rendered as intro + clip + outro in a single encode, fitted to vertical over a blurred background.
• Per project and per clip – a project follows the default intro/outro, uses none or a specific item; each clip can follow the project, use none or pick its own. An unusable asset is skipped with a plain-language event instead of failing the clip.
• Layers – one timeline lane for image/video overlays, text and sound (update_clip { layers }, reelsaw clip-layers): position, size, opacity, fonts shipped in the image, several music tracks with fade, loop and sidechain ducking under speech. Right-click a layer on the 9:16 preview to edit, centre, reorder, copy or remove; double-click text to retype it. ⌘C / ⌘V copy and paste layers.
• Layer templates – save a look once (a logo pop in the first seconds, a follow card at the end, a watermark or a music bed, plus an intro and outro) with timing from the start, the end, the whole clip or a share of its length, then apply it to selected clips or set it on a project so every new clip gets it (list_layer_templates, apply_layer_template, reelsaw templates).

Agents and automation

• MCP server – /api/mcp (Streamable HTTP) lets Claude Code, Claude Desktop, Cursor or a script create projects, wait for them, read timed transcripts, edit layers and framing, create and re-render clips, manage the library and layer templates, and export the results. On a hosted plan the same tools are available, including Free. Authenticate with an API key from Settings, or with MCP_OAUTH=true so claude.ai and Claude Desktop connectors sign in through the browser.
• Operator CLI – docker compose exec app reelsaw … inspects and fixes a running instance: status, events, clip, rerender, stuck, recover, logs, cuts, align, hear, flags, storage, retention, deleted, restore, purge, users, approve, make-admin, set-password, library, templates, apply-framing, import-report and more (every command is in Operator CLI).

Hosting and accounts

• Accounts, approval and quotas – the first account is the admin; SIGNUP_MODE=open|approval|closed decides who gets in, and admins approve, suspend, promote and remove people on the Users page, with usage per account (projects, source minutes, clips, model calls and tokens, stored bytes). Optional per-user quotas, retention of old files, captcha on sign-in, email (approval notices, password reset, verification) and terms/privacy links. Everyone has their own Project defaults (vocabulary, instructions, language, clip length, pauses, caption style, intro and outro) and API keys.
• Models and providers at runtime – admins switch the language, transcription and alignment models under Settings → System → Models without a restart: pick a provider, enter its key or server URL there (stored encrypted, never shown again), and choose from live model lists; local Ollama and Speaches models download with a progress bar. Keys and models are tested when saved.

Operations

• Reliable, resumable, scalable – every stage is a pg-boss job in Postgres that checkpoints its output, so a crashed or restarted worker picks up where it left off; a worker that stops heartbeating (every 15 s) has its job re-queued by pg-boss within about a minute or two, and a reconciler re-enqueues anything orphaned. A long stage saves its parts as they finish (each 3-minute window of Finding moments), so a retry or a timeout on a slow local model continues where it stopped instead of starting over, and a timeout says how far it got (LLM_CALL_TIMEOUT_SECONDS, STAGE_IDLE_TIMEOUT_MINUTES). Run extra worker containers to scale. Every run (first pass, retry, re-render, re-analyze) is recorded with its own timings, and the UI shows elapsed against expected time per stage, learned from your finished projects. The browser gets live updates over a tRPC SSE subscription with automatic reconnect.
• Storage and install – local volume or s3 (AWS S3, MinIO, Cloudflare R2, Backblaze B2…); superseded renders, expired sources and orphaned files are swept hourly (reelsaw storage reports, reelsaw storage sweep runs it). One-line installers for macOS/Linux (bash) and Windows (PowerShell), and Report an issue saves a zip of a project's data, decisions and logs, with secrets left out.
• Observable and fail-fast – every service is built in main and self-checks, so the process exits with a readable reason if a model id is wrong, a key is missing, a binary is absent, the bucket is unreachable or the DB is down. LLM_TRACE=logs or langfuse captures every prompt/response (tokens, latency, project/clip/stage), and OTEL_ENABLED=true turns on OTLP traces and metrics for HTTP, Postgres, pg-boss jobs, AI SDK calls, ffmpeg/yt-dlp and each stage, with stage, download and transcription time per second of video and per-proxy download health.

Contributing

See CONTRIBUTING.md (principles, tests, sign-off). Roadmap for learning from user decisions: docs/SELF-IMPROVEMENT.md.

Read docs/PRINCIPLES.md first: the consistency, concurrency and UI rules every change follows. Overlay stacking (Select inside a Dialog, z-index, portals) and the mobile viewport (h-dvh, not h-screen) are docs/UI.md. Agent instructions live in AGENTS.md (the only agent file; see the process principles).

Install (no source needed)

Needs Docker (Docker Desktop on macOS and Windows). One command: it looks at the machine (RAM, GPU, Docker's memory, free disk, an Ollama already running and the models it has, the port), proposes a setup, asks three questions in a terminal (local Ollama or a hosted provider, which model, go ahead) and installs Ollama for you (Homebrew, ollama.com or winget) if you pick local and it is missing:

curl -fsSL https://raw.githubusercontent.com/osbytes/reelsaw/main/deploy/install.sh | bash

On Windows (Docker Desktop), in PowerShell:

irm https://raw.githubusercontent.com/osbytes/reelsaw/main/deploy/install.ps1 | iex

WSL2 users can run the bash one-liner as is (with Docker Desktop's WSL integration on).

It writes ~/reelsaw (compose file + .env with generated secrets), pulls the published image and the model, starts everything and waits until http://localhost:3000 answers (or the next free port). Fully local is the default when the machine can carry a useful model: qwen2.5:14b (~9 GB) with 24 GB of RAM or a 12 GB GPU, qwen2.5:7b from 12 GB, gemma3:4b below that, and a good model that Ollama already has is reused instead of downloaded; with an NVIDIA GPU and the container toolkit, Ollama and Speaches run on the GPU (and renders encode with NVENC; see "Using a GPU" below). Hosted providers (OpenAI, Anthropic, Google, Groq, Vercel AI Gateway) are a choice in the same prompt: the key is asked for and never echoed, and Anthropic and Google pair with local Speaches for speech-to-text. Without a terminal, or with REELSAW_YES=1, the proposed answers are taken; REELSAW_AI, REELSAW_LLM, REELSAW_PORT and friends answer in advance and REELSAW_DRY_RUN=1 prints the plan and stops (every option is listed at the top of the script). Manual setup: deploy/README.md. The scripts are deploy/install.sh and deploy/install.ps1; they are served from main (see docs/RELEASING.md).

Quick start (from source)

git clone <this repo> reelsaw && cd reelsaw
cp .env.example .env
# edit .env: set BETTER_AUTH_SECRET, APP_URL, and at least one AI provider key
docker compose up -d --build

Open http://localhost:3000, create an account (it becomes the admin), start a project. Set SIGNUP_MODE=closed or approval afterwards if the instance is public.

Scale workers:

docker compose --profile workers up -d --scale worker=2

Each worker keeps its own scratch copies of the sources and renders it works on. With an S3 store they are a cache: idle copies of finished projects are evicted after SCRATCH_CACHE_HOURS (24), a job short of disk evicts the least recently used ones first, and a worker that still cannot make room hands the job to a worker that has it and stops taking media jobs until its disk recovers (Settings → System → Workers shows "disk full"). A job leases the scratch it is about to write against its volume's live leases before the first byte lands, so two large ingests on one box cannot both pass the free-space check and both run out mid-download: the second waits for the first. Processes sharing a volume (the app and worker replicas of one host) share one ledger; a worker on another host never contends.

Use MinIO for local S3 testing:

docker compose --profile minio up -d
# then in .env: STORAGE_DRIVER=s3 S3_ENDPOINT=http://minio:9000 S3_FORCE_PATH_STYLE=true S3_BUCKET=reelsaw S3_ACCESS_KEY_ID=reelsaw S3_SECRET_ACCESS_KEY=reelsaw123

(Create the bucket once via the MinIO console at http://localhost:9001.)

Model configuration

The language model and the transcription model are independent; mix vendors freely (Anthropic has no speech-to-text, so pair it with Deepgram/OpenAI/AssemblyAI/Groq). Each needs only its own key.

The transcription model must return word timestamps (captions and cuts need them), so openai:gpt-4o-transcribe, which returns none, is not supported.

Fully local (no API keys)

Ollama only serves language models; it has no speech-to-text. So the local stack is Ollama for AI_MODEL plus Speaches (faster-whisper behind an OpenAI-compatible API, word timestamps supported) for TRANSCRIPTION_MODEL. Both ship as a compose profile:

docker compose --profile local-ai up -d            # Speaches; use your native Ollama (macOS: Metal GPU)
docker compose --profile local-ai --profile ollama up -d   # Linux: also run Ollama in a container
ollama pull qwen2.5:14b                            # or: docker compose exec ollama ollama pull qwen2.5:14b

.env:

AI_MODEL=compat:qwen2.5:14b            # qwen2.5:7b or gemma3:4b on less RAM
OPENAI_COMPATIBLE_BASE_URL=http://host.docker.internal:11434/v1   # native Ollama; http://ollama:11434/v1 for the container
OPENAI_COMPATIBLE_API_KEY=ollama
TRANSCRIPTION_MODEL=openai:Systran/faster-whisper-small
OPENAI_BASE_URL=http://speaches:8000/v1
OPENAI_API_KEY=local

Notes: the app requests response_format: json_schema from the compat server (OPENAI_COMPATIBLE_STRUCTURED_OUTPUTS=false to disable) and falls back to lenient parsing when a model ignores it, but small models still pick weak moments and write bland hooks. gemma3 4b works; qwen2.5 14b (the installer's default on 24 GB or more), llama3.1 8b or gemma3 12b are noticeably better. On first boot the app auto-downloads the whisper model into Speaches and verifies the Ollama model is pulled (it exits with ollama pull <model> otherwise). Speaches CPU with faster-whisper-small runs roughly real-time; use the latest-cuda image and faster-whisper-large-v3 on a GPU. The OPENAI_* variables are reused for Speaches because it speaks the OpenAI audio API, which is why the LLM goes through the compat provider instead.

Using a GPU

Three things can use one, each in its own container, so the GPU is a Compose override rather than a different image: docker-compose.gpu.yml reserves every NVIDIA device for speaches, ollama, app and worker. Host side you need the NVIDIA driver (nvidia-smi works) and the NVIDIA Container Toolkit (sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker; on Windows, Docker Desktop on WSL2 with the Windows driver only). Then in .env:

# Linux/macOS (colon). On Windows use a semicolon: docker-compose.yml;docker-compose.gpu.yml
COMPOSE_FILE=docker-compose.yml:docker-compose.gpu.yml
SPEACHES_IMAGE=ghcr.io/speaches-ai/speaches:latest-cuda
WHISPER_MODEL=Systran/faster-whisper-large-v3      # or deepdml/faster-whisper-large-v3-turbo-ct2
TRANSCRIPTION_MODEL=openai:Systran/faster-whisper-large-v3

and docker compose --profile local-ai up -d as before. What changes:

• Transcription is the big one: faster-whisper on CUDA runs large-v3 in a fraction of real time, where small on a CPU was about real time (the transcribe factor in the table below).
• The language model uses the card when Ollama runs in the ollama container (--profile ollama). A native Ollama (macOS, or Linux on the host) already uses the GPU on its own.
• Renders encode with h264_nvenc instead of libx264: FFMPEG_ENCODER=auto probes it at boot and keeps libx264 when the card is not there. Filters (crops, blur, captions) still run on the CPU, and render was already the short stage, so expect seconds, not minutes. A hardware encode that fails mid-clip (NVENC session limit, driver reset) is redone with libx264 for that clip.
• The on-device models (face, text, alignment) stay on the CPU by default. ONNX_EXECUTION_PROVIDER=cuda moves them, but the published image carries no CUDA libraries for onnxruntime, and on Apple silicon CoreML measured slower than the CPU for all three; see docs/CONFIG.md.

Settings → System → Overview and docker compose exec app reelsaw machine show the GPU the app process sees, the encoder in use and why, and where each on-device model runs. Workers (and reelsaw workers) lists every API and worker replica with its specs and what it is processing — needed when the HTTP process is not the one doing the work. macOS has no CUDA in Docker: run Ollama natively for Metal, and leave Speaches on the CPU (or point OPENAI_BASE_URL at a whisper.cpp or MLX server running on the host, which speaks the same API).

What to expect: processing time

Every stage scales with the length of the video. Expected seconds for a stage ≈ video seconds × factor, where the factor depends on the models (local vs vendor) and your machine:

Rule of thumb: a vendor stack finishes a video in about a fifth of its length; a local CPU stack takes about as long as the video. Refine runs alongside finding moments, so the stages don't add up exactly. Finding moments saves each part it finishes, so on a long video and a slow model a retry or a timeout continues from the last finished part instead of starting over, and the project shows "N of M parts, about X left".

¹ Apple-silicon laptop, qwen2.5:14b on Ollama + faster-whisper-small on Speaches (CPU), 20-minute podcast: download 15 s, transcribe 4.9 min, refine 10 min, find moments 4 min, render 47 s; 20 min in all. On a local 14B model the disfluency verifier (one call per candidate, ~40 s each) is the bottleneck, not transcription; a smaller model or DISFLUENCY_LLM=false cuts refine to a fraction. ² Same video through the Vercel AI Gateway, whisper-1 + claude-haiku-4.5: download 13 s, transcribe 77 s, refine 31 s, find moments 24 s, render 56 s; 3.4 min in all. Gemini 2.5 Flash took 5 min (long "thinking" outputs), Sonnet 4.5 about 4. Mixing a vendor LLM with a local TRANSCRIPT_ALIGN_MODEL makes render the slowest stage (5–9 min: every cut is re-aligned on the CPU); keep alignment on the same vendor as transcription. Full numbers in eval/bench/results/.

These factors are only the starting point. The app learns from its own finished projects: each stage's expected time is the median over your recent projects that ran with the same models, shown next to the elapsed time on the project page while it runs, in the activity feed when a stage finishes (Transcribe finished in 4:50 (expected ~4:00)), and as a per-stage expected vs actual table under Settings → System → Machine. There is no calibration step; the runs are the calibration.

Switching models without a restart

.env names the default models and keys; an admin can change both at any time under Settings → System → Models. Each of the three pickers (language, transcription, alignment) opens provider first: pick a provider and, if it has no key or server URL yet, enter it right there; its models then appear. Cloud vendors are asked live, a self-hosted Ollama or Speaches shows what is installed and downloads anything else with a progress bar, and the search box filters every provider at once. Saving a key tries it first (the provider's models are listed with it) and a wrong one is refused; saving a model checks it too (key present, server reachable and model installed, transcription returns word timestamps). Keys are stored encrypted in the database with a key derived from BETTER_AUTH_SECRET and never shown again (the UI sees the last four characters); rotating that secret makes stored keys unreadable, and the UI asks for them again. Changes survive restarts and reach worker processes within seconds; a stage already running finishes with the model it started with. "Back to .env" / "Remove" drop an override. Over the API: admin.models.*, admin.providers.*.

Accounts and roles

The first account created on an instance is the admin; everyone else is a user. Admins see the system settings (models, storage, resource meters) and the Users page; users see their own projects and settings only. SIGNUP_MODE decides who gets in:

Every user has Settings → Project defaults (vocabulary, instructions, language, clip length, pauses, caption style…) applied to each new project they or their agents start; the new-project form starts from them and can save its current values back. Admins also see the System section (models, storage, resource meters) and /api/health includes machine numbers only for them (or with HEALTH_DETAILS=true).

Suspending an account (Users page) ends its sessions and refuses its API keys at the next request; removing one deletes its projects, clips and files. Locked out of the password? Sign-in has Forgot password when email is on. Without email, an operator sets a new one from the shell (docker compose exec app reelsaw set-password you@example.com); --must-change makes them pick their own at the next sign-in, without asking for the temporary password (the session is the proof they knew it). To promote an account from the shell: docker compose exec app reelsaw make-admin you@example.com.

Email (EMAIL_DRIVER=smtp|resend, EMAIL_FROM, SMTP_URL or RESEND_API_KEY) turns on approval notices, password reset and, with REQUIRE_EMAIL_VERIFICATION=true, verification before the first sign-in. Confirmation and reset links work for 24 hours and only once; an expired link lands on a page that sends a new one. Sign-in has Forgot password? when email is on; choosing a new password signs out other devices. Without email everything else works; approvals are handled on the Users page.

Hosting for many people

Everything below is off by default and switched on by environment (deploy/.env.example lists them), so a single-user box stays as it is:

• Quotas: MAX_PROJECTS_PER_USER, MAX_SOURCE_MINUTES_PER_DAY, MAX_SOURCE_MINUTES_PER_MONTH, MAX_NEW_PROJECTS_PER_HOUR, MAX_STORAGE_MB, MAX_LIVE_JOBS_PER_USER (jobs queued or running from one account's own requests), MAX_RERENDERS_PER_HOUR (re-renders, hand-made clips and re-runs such as re-analyze and refine, per hour), MAX_REQUESTS_PER_MINUTE and MAX_REQUESTS_PER_HOUR (0 = unlimited), with per-user overrides on the Users page. Storage counts sources, uploads (an upload already copied into a project counts once), rendered clips and intros/outros, and is checked on every way a file arrives. A refusal is a sentence that says what to do ("Monthly limit of 300 minutes reached (312 used); resets on 1 November"). The monthly window is the account's billing period when it has a plan, else the UTC calendar month. On an account with a minutes limit, a link whose site reports its length is refused when the project is created (with the upgrade offer), so no project is made for it. Otherwise the length is checked again before the download and once the file is probed, so an over-long video fails the project with that reason, once, without being fetched again. The first video of a plan with a free allowance is the first one whose length is known.
• Watermark: WATERMARK_NON_ADMINS (default on) burns the logo (WATERMARK_LOGO_PATH) with WATERMARK_TEXT ("reelsaw.com") under it into the clips of accounts that are not admins; a paid plan removes it, and an admin can switch it per user on the Users page (Limits). Decided at render time, so a re-render after the change applies it. Clips already rendered keep what they were rendered with.
• Usage (Users → Usage): projects, minutes of source, clips rendered, model calls and tokens (every model call writes a llm_usage row attributed to the project's owner), stored bytes.
• Retention: SOURCE_RETENTION_DAYS / CLIP_RETENTION_DAYS delete stored files from finished projects (the store object and the local copy); rows and history stay. A clip re-renders from its source; once the source is removed, "fetch again" (Retry on the project) downloads it first. reelsaw retention runs a sweep by hand. With an S3 store, SCRATCH_CACHE_HOURS (24) is how long a worker keeps its local copies of stored files once a project is finished and idle; the bucket copy stays and comes back when a job needs it.
• History pruning: EVENT_RETENTION_DAYS (90), LLM_USAGE_RETENTION_DAYS (180) and ADMIN_EVENT_RETENTION_DAYS (365) delete old rows in batches once a day (0 keeps everything); a project's run skeleton (created, run started, done) is always kept.
• Bot protection: CAPTCHA_PROVIDER (Cloudflare Turnstile, reCAPTCHA, hCaptcha) with site and secret keys puts a widget on sign-up, sign-in and password reset.
• Free-tier abuse (all off unless set; apps/server/src/abuse.ts): MAX_SIGNUPS_PER_IP_PER_DAY caps accounts per network address (behind a proxy, TRUST_PROXY reads CF-Connecting-IP, X-Real-IP, X-Forwarded-For); BLOCK_DISPOSABLE_EMAILS refuses temporary-address domains; one account per real mailbox (john.doe+x@gmail.com is johndoe@gmail.com); MAX_REQUESTS_PER_MINUTE / _HOUR cap changes (API mutations, MCP tool calls) per account, with plan and per-user overrides like the other quotas. An hourly pass flags accounts that share a sign-up address, arrive in bursts or use a temporary domain; the Users page shows the flags ("Flagged only"), reelsaw abuse lists them, and nothing is suspended automatically. Every refusal is a sentence.
• yt-dlp from a datacentre: YouTube refuses many cloud IPs. YTDLP_PROXY is one URL or an operator-ranked pool (a home Pi, a cheap VPS, a paid proxy). {session} in a URL is replaced per download so a residential gateway can sticky-session one video and rotate the next. Reliability is a circuit breaker on each candidate and never changes the ranking among healthy ones. YTDLP_COOKIES_FILE helps too; a blocked download fails with "upload the file instead". reelsaw ytdlp-proxies shows the pool.
• Legal: TERMS_URL / PRIVACY_URL appear on the sign-up form.
• MCP over OAuth (MCP_OAUTH=true): Reelsaw becomes an OAuth 2.1 authorization server for MCP connectors (claude.ai, Claude Desktop): dynamic client registration, sign-in and a consent page in the browser, discovery at /.well-known/oauth-protected-resource/api/mcp (which names the issuer, <APP_URL>/api/auth, whose metadata is at /.well-known/oauth-authorization-server/api/auth). API keys keep working for CLIs. The consent page shows the address the approval returns to, registrations are capped at 20 per network address per day, and each person lists and revokes the apps they allowed under Settings → Account → Connected apps.
• Client address behind a proxy (TRUST_PROXY, TRUSTED_PROXY_HEADER, TRUSTED_PROXY_HOPS): the sign-up cap and the registration limit key on the client address. With the default x-forwarded-for it is the entry TRUSTED_PROXY_HOPS places from the right (1 behind Caddy, 2 behind Cloudflare and Caddy), so addresses a client prepends do not count; the other headers are read only when selected.

Billing (enterprise edition)

The hosted service at app.reelsaw.com adds plans and billing through Stripe. That code lives in apps/server/src/ee and apps/web/src/ee under its own licence (LICENSING.md), ships in the same image, and runs only with REELSAW_EE=true plus STRIPE_SECRET_KEY and STRIPE_WEBHOOK_SECRET. Nothing else changes for a self-hosted install.

• A plan is a named set of the limits above (ee/billing/plans.ts); the account's plan sits between the instance setting and the admin's per-user override. Free is a plan too: what an account with no, or a lapsed, subscription gets.
• Stripe is the source of truth. A webhook (POST /api/billing/webhook) is acknowledged only once its event row and a billing.sync job have committed together; the job rewrites the account's row from Stripe, so events that arrive late, twice or out of order converge. A daily reconcile and reelsaw billing sync repair anything a lost webhook left behind.
• Settings → Billing shows the plan, the period's usage and the plan grid; Checkout and the Billing Portal are Stripe's pages. get_billing, billing_checkout_url and billing_portal_url do the same over MCP; reelsaw billing <email> shows the row, Stripe's view and the last webhook events.
• A failed payment keeps the plan for BILLING_GRACE_DAYS (7), then the account is on Free. Admins can grant a plan without Stripe (Users → Limits → Plan, or reelsaw billing set-plan). A lapse or downgrade does not re-watermark clips already rendered; only new renders get the watermark. An account that already has a live Stripe subscription is sent to the Billing Portal to change it, never to a second Checkout.
• Setting up a Stripe account: pnpm --filter @reelsaw/server stripe:setup --webhook --portal (rebuilds @reelsaw/shared first) creates the products, prices (by lookup key), webhook endpoint and portal configuration from the plan list, idempotently. Developing without a Stripe account: pnpm --filter @reelsaw/server stripe:mock stands in for api.stripe.com (STRIPE_API_BASE=http://host.docker.internal:4242) and test/billing-stack.test.ts drives the whole lifecycle through signed webhooks.

Design and decisions: docs/specs/ee.md.

MCP: drive Reelsaw from an agent

Reelsaw is an MCP server (Model Context Protocol, Streamable HTTP) at /api/mcp. Create an API key in Settings → API keys and MCP (shown once) and point a client at it:

claude mcp add --transport http reelsaw https://your-host/api/mcp --header "Authorization: Bearer rs_…"

{ "mcpServers": { "reelsaw": { "type": "http", "url": "https://your-host/api/mcp", "headers": { "Authorization": "Bearer rs_…" } } } }

Key permissions. A key acts as you, with one exception: the instance-administration tools (get_instance_settings, set_instance_settings, list_workers, get_queue, list_failures, get_failure, retry_failures, check_for_update, and instance-wide report_issue) need a key created with Allow instance administration (admin tools over MCP), a checkbox admins see in Settings. Keys are created without it by default, and a key that lacks it is refused with "This key cannot change instance settings". Keys created before permissions existed were given the administration grant when the database upgraded, so they keep working; it shows on the key in Settings, where you can revoke it or replace the key. A key with no permission record is never an administrator. Connectors that sign in through OAuth (MCP_OAUTH=true) never get the admin tools, whoever signs in.

Clients that launch a command instead of calling a URL with a header (Claude Desktop without MCP_OAUTH, Cursor's stdio entries) use reelsaw mcp: the same tools over stdio, forwarded to a running instance. It needs no database or models, only REELSAW_URL (default: the app in the same container) and REELSAW_API_KEY:

{ "mcpServers": { "reelsaw": { "command": "docker", "args": ["exec", "-i", "-e", "REELSAW_API_KEY=rs_…", "reelsaw-app-1", "reelsaw", "mcp"] } } }

Reelsaw is listed on the MCP registry as io.github.osbytes/reelsaw (server.json, published on every release), so registry-aware clients can install the bridge from the image without a running container to exec into:

{ "mcpServers": { "reelsaw": { "command": "docker", "args": ["run", "-i", "--rm", "-e", "REELSAW_URL=https://your-host", "-e", "REELSAW_API_KEY=rs_…", "ghcr.io/osbytes/reelsaw", "reelsaw", "mcp"] } } }

Automation tools can call the same endpoint without an MCP client: it is stateless and answers plain JSON, so a tools/call POST is one HTTP request. docs/n8n has a ready-made n8n workflow that clips every new upload on a YouTube channel.

The tools (generated from apps/server/src/mcp-tools.ts; each tool's full description and input schema say what it returns, whether it runs in the background and which neighbour to use instead):

get_project / get_clip show each clip's effective intro/outro and where it comes from. They call the same tRPC procedures as the web app, so ownership checks and validation are identical (apps/server/src/mcp.ts; the catalogue both transports advertise is apps/server/src/mcp-tools.ts). The same key also authenticates tRPC (/api/trpc/*), uploads and media downloads, so an agent can fetch the rendered clips. Keys are hashed at rest (@better-auth/api-key); revoke them from Settings.

Operator CLI

docker compose exec app reelsaw <command> (see TROUBLESHOOTING for when to reach for which). Generated from the header of apps/server/src/cli.ts.

Run as docker compose exec app reelsaw <command> (from source: pnpm --filter @reelsaw/server cli <command>).

Settings and precedence

Where a value comes from, from the weakest source to the strongest (a later link overrides an earlier one):

Two things to know:

• An env / instance refine gate wins. DISFLUENCY_REMOVAL=false (or the matching Pipeline setting) skips filler removal for everyone even when a project asks for it; the new-project form disables that toggle and says so. TRANSCRIPT_BIAS=false still silently ignores vocabulary hints (.env only). If a project setting seems to do nothing, check the gate first (TROUBLESHOOTING, "A project or clip setting has no effect").
• Existing projects keep their frozen settings. Changing your Project defaults affects new projects only.

Code: apps/server/src/pipeline/settings.ts (project), resolveStyle in packages/shared/src/index.ts (style), apps/server/src/pipeline/render.ts (framing), apps/server/src/ai/models.ts (models), apps/server/src/quota.ts (quotas).

Architecture

apps/web      React 19 + Vite + Tailwind + TanStack Query + tRPC client + Better Auth client
apps/server   Hono + tRPC + Better Auth + Drizzle (Postgres) + pg-boss workers + ffmpeg/yt-dlp
packages/shared  zod schemas and types shared by both

Server composition (apps/server/src/services.ts): createServices() builds config → media tools → ai → storage → database (+migrations) → events → queue → auth, checking each one. The resulting Services object is injected into the tRPC context, the HTTP app and every pipeline stage (StageContext), so nothing reaches for a global. Boot logs one boot step ok line per step (JSON, or pretty on a terminal) with the step name and its duration in ms: config, data dir, resources, media tools, LLM call capture, AI models, the on-device detectors, storage, database, model settings, event bus, job queue, email, auth. Steps whose feature is switched off say so in their name (face detector (disabled)). A failing step stops the process with exit code 78 and a message that names the step and the fix, for example a transcription model that returns no word timestamps or a model server that is unreachable. Some steps depend on the process's role: a process with RUN_WORKERS=false only warns when free disk is under MIN_FREE_DISK_MB (so a full disk cannot lock out the admin UI that frees it) and skips the face, text and voice detectors; a process with RUN_HTTP=false skips the mail server check. The AI-models step checks the environment's models; once the database is up the effective models (after an admin's runtime override) are checked too.

Pipeline (each box is a pg-boss queue; each stage emits events and enqueues the next):

project.ingest ─▶ project.transcribe ─┬▶ refine.entities   ─┐
 yt-dlp / upload   audio → chunks       ├▶ refine.disfluency ─┼▶ project.analyze ─▶ clip.render (×N)
 ffprobe           parallel transcribe  └▶ refine.scenes     ─┘  LLM picks moments    layout + captions
 store source      checkpoint per chunk    (fan-out, fan-in       snap to sentences    cuts + crossfades
                                            on the transcript row)                     finalize project

The three refine jobs are independent: each writes its result under a row lock on the transcript (entity flags win over overlapping disfluency flags), and the last one to finish enqueues analyze. A refine step that exhausts its retries is marked failed and skipped; it never blocks the project. "Re-run refine" re-queues any subset on the existing transcript, e.g. after editing the vocabulary. Chunk transcription runs TRANSCRIBE_CONCURRENCY calls in parallel (auto: 1 for a self-hosted whisper, 3 for cloud vendors).

Durability model:

• pgboss.job is the source of truth for "what should run". projects.status / clips.status is the source of truth for "what happened".
• Each stage checks for existing output first (source file in storage, transcript chunk rows, clip rows, rendered clip in storage) and skips work that is done.
• Queues use policy: exclusive + singletonKey so a stage can never run twice concurrently for the same project/clip.
• heartbeatSeconds (15 s) on every queue: if a worker dies, pg-boss notices the missed heartbeat on its next supervisor pass (every 60 s), fails the job and retries it elsewhere after a 15 s delay that doubles per retry. Budget a minute or two, not seconds; each interruption uses one retry.
• On boot and every 5 minutes a reconciler re-enqueues any in-progress project or clip that has no live job.
• Manual Retry resumes from the furthest completed stage; Re-analyze keeps the transcript and picks new clips.

Development

Live-reload loop in Docker (source mounted, no image rebuilds; rebuild the dev image only when dependencies change):

docker compose -f docker-compose.yml -f docker-compose.dev.yml --profile local-ai up -d app
open http://localhost:5173      # Vite + HMR; /api and /media proxied to the server (tsx watch)
docker compose exec app reelsaw stuck

Models for running outside Docker: pnpm models:download (face, on-screen text, voice activity, forced alignment; ~105 MB).

pnpm install
docker compose up -d db
cp .env.example .env   # set DATABASE_URL=postgres://reelsaw:reelsaw@localhost:5432/reelsaw
pnpm dev               # server on :3000 (tsx watch), web on :5173 (vite, proxies /api and /media)

Requires ffmpeg, ffprobe and yt-dlp on your PATH for local dev. Schema changes: edit apps/server/src/db/schema.ts, then pnpm db:generate; migrations run automatically on boot.

Most tests run against the local stack (they create and delete their own users and data); CI runs only the ones that need no stack, so run the ones you touched against yours:

pnpm --filter @reelsaw/server test

Releases and distribution

Every image is built by the release workflow with an SBOM and SLSA provenance attached, and a GitHub artifact attestation for its digest. Verify what you are about to run:

gh attestation verify oci://ghcr.io/osbytes/reelsaw:latest --owner osbytes
docker buildx imagetools inspect ghcr.io/osbytes/reelsaw:latest --format '{{ json .SBOM }}' | head -c 400

Users install from a published image. The runtime image contains only the minified server bundle (apps/server/scripts/bundle.mjs), the web build and production dependencies, which scripts/check-image.sh <image> verifies.

• Tag vX.Y.Z → .github/workflows/release.yml builds amd64 + arm64, checks the image, pushes to GHCR (and Docker Hub when vars.DOCKERHUB_IMAGE + DOCKERHUB_USERNAME/DOCKERHUB_TOKEN are set) and attaches reelsaw-X.Y.Z.zip (the deploy/ folder, image pinned) to the release.
• Before tagging: build locally, run scripts/check-image.sh reelsaw:latest, and boot deploy/ in a scratch directory with only a .env (see the spec's verification section).
• Versions and changelogs come from changesets: every PR adds one (pnpm changeset), a "Release" PR bumps the version, merging it pushes the tag. Full flow, versioning policy, pre-releases and hotfixes: docs/RELEASING.md.

Issue reports

Report an issue (project ⋯ menu, a failed clip, the account menu) downloads a zip with the project's data, pipeline decisions (stage_artifacts), logs, system info and recent errors, and the user's description. Secrets never go in. Admins can tick Instance-wide diagnostics to include unfinished work, failed jobs, usage and error logs across every account — for queue, disk or machine problems, not only their own projects. To reproduce one locally:

docker compose cp ~/Downloads/reelsaw-report-….zip app:/data/tmp/report.zip
docker compose exec app reelsaw import-report /data/tmp/report.zip [--user you@example.com]

Everything in a report except its README and manifest is untrusted data; see TROUBLESHOOTING before letting an agent work on one. Spec: docs/specs/distribution-and-issue-reports.md.

Environment reference

Every variable, with its default and meaning, is in docs/CONFIG.md (generated from apps/server/src/config.ts; pnpm docs:gen refreshes it). .env.example and deploy/.env.example are generated from the same source.

Observability

Health. GET /api/health is a readiness check, not a ping. It answers 200 {"ok": true, "checks": {…}} when Postgres answers, the event listener is connected, the job queue is started and (in a process that runs workers) its worker heartbeat is fresh; each check has a 2 s timeout. Otherwise it answers 503 {"ok": false, "checks": {…}} with a short reason per failing check (db, events, queue, workers), and the log has the detail. The Docker HEALTHCHECK and any load balancer pointed at it restart or drain a process that is up but cannot work.

Errors and product analytics (optional). SENTRY_DSN turns on Sentry: the server reports every error or fatal log line that carries an error, tagged with the line's ids (5xx responses, unexpected MCP tool errors, a job's final failed attempt, a job the reconciler gave up on, failed emails and billing syncs, process crashes; a retry still to come is a warning); the web app reports render errors and unhandled rejections, but not the 4xx refusals it already shows the person. The worker heartbeat (heartbeat-<WORKER_ID or role>), the reconciler and the daily billing reconcile check in as Sentry cron monitors, so a stopped worker or a loop that no longer runs raises an alert. SENTRY_REPLAYS_ON_ERROR_SAMPLE_RATE (default 0) keeps a fully masked replay of the moments before a browser error. POSTHOG_KEY (and POSTHOG_HOST, EU cloud or self-hosted) turns on PostHog: the server sends sign-ups, approvals, email verification, project created / done / failed / cancelled, clip done / failed, jobs a person asked for, quota refusals, clip downloads and exports, MCP tool calls (tool, outcome, client name), and billing and abuse-flag events, each with the surface it came through (web, api, mcp, cli); the web app sends page views and a handful of product events (lib/telemetry.ts). Event properties are ids, counts, durations and names, never a title, an error message or other free text; people are identified by user id, role and plan only, with no autocapture and no session recording. Both are off unless set, the browser gets the public keys from system.status, and the SDK chunks are only requested when on. Prompts, transcripts and media never go through either (apps/server/src/telemetry.ts). Release images upload their source maps to Sentry when the release workflow has SENTRY_AUTH_TOKEN (.github/workflows/release.yml); no map ships in the image.

OTEL_ENABLED=true
OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318   # OTLP/HTTP; traces + metrics
OTEL_SERVICE_NAME=reelsaw

Sentry's OTLP endpoint takes traces and logs (…/integration/otlp, plus OTEL_EXPORTER_OTLP_HEADERS). Each pino line that passes its level is exported to /v1/logs, with the active trace id attached; prompt, transcript and address fields stay in the local log. Sentry does not ingest OTLP metrics. With SENTRY_DSN set and OTEL_METRICS_ENABLED left on, the server sends Node runtime gauges (memory, CPU, event loop) and the metrics below through the DSN instead. Set OTEL_METRICS_ENABLED=false to keep traces and logs and drop those metrics.

Metrics. Durations are seconds. Attributes are bounded sets (stage, proxy id, model, outcome), never a project, clip or user id. Work grows with the video, so the stage, download and transcription numbers each come with a realtime factor: wall seconds per second of media (0.25 = a 1-hour video takes 15 minutes), comparable across a 30-second clip and a 2-hour talk.

Over OTLP, the auto-instrumentation adds http.server.request.duration, and pg-boss adds messaging.process.duration and pgboss.queue.jobs (backlog by queue and state). A proxy dashboard: reelsaw.ytdlp.download.realtime_factor p50/p95 by proxy, reelsaw.ytdlp.attempts rate by proxy and outcome, and reelsaw.ytdlp.proxy.healthy by proxy.

Spans: inbound HTTP named by route (POST /api/trpc/projects.create; /api/health excluded), Postgres queries, outbound fetch to AI vendors, pipeline.<stage> per job with reelsaw.project_id / reelsaw.clip_id / reelsaw.attempt, exec ffmpeg|ffprobe|yt-dlp, and the AI SDK's own ai.generateText / ai.transcribe spans (inputs/outputs not recorded). pg-boss propagates trace context from send() into the worker span, so one trace follows a project from the create mutation through every stage. A client span with no parent (an S3 read or a query no request or job is around, such as queue polling) is not recorded unless OTEL_TRACES_SAMPLER is set.

The browser exports its own spans as reelsaw-web (page load, navigation, fetch, and live subscriptions; not clicks) and sends traceparent on API calls, including the subscription stream, so a click and the server work it starts are one trace. OTEL_TRACES_SAMPLE_RATE (default 1) is the fraction kept, on the browser and the server together. The Sentry browser SDK does not trace; those spans would be a second copy.

Entity-aware transcription

Whisper mishears names ("Copia" → "Copious"). Three layers fix that, all optional:

1. Vocabulary hints (TRANSCRIPT_BIAS=true): proper nouns pulled from the channel name, title, description, tags and chapters, plus anything typed into "Names & terms" on the project form, are passed to the transcriber (prompt for OpenAI/Groq/Speaches, keyterm for Deepgram, wordBoost for AssemblyAI). Biggest win: in our eval it halves WER and lifts entity recall from 64% to 95% at zero cost.
2. Constrained LLM correction (TRANSCRIPT_CORRECTION=true): transcript spans that sound like a known entity become candidates; the language model confirms each one in context with a confidence score. Only candidates can change, so the model cannot rewrite the transcript, and word timestamps are kept.
3. Review in the UI: corrections at or above TRANSCRIPT_AUTO_FIX_CONFIDENCE (0.85) are applied automatically and shown as "auto-fixed"; the rest appear under the transcript line with the confidence and accept/reject buttons. Applying decisions re-renders only the clips whose captions overlap the change.

Language: the filler/repeat rules, the sound-alike name search and the CTC cut aligner are English-only. For a transcript detected as any other language those steps are skipped with a plain-language event ("not available for de") instead of running blind; the language-model checks and the whisper-based cut search still work.

Disfluencies

Fillers ("um"), repetitions ("the the") and false starts ("your pay actual pay dates") are detected after transcription (DISFLUENCY_REMOVAL=true): a rule pass with near-certain confidence, then a model pass (DISFLUENCY_LLM=true) that rewrites each window cleanly and is diffed against the original so it can only ever delete. A model deletion is kept only when every removed word is a filler or is restated within the next few words; repairs land as pending flags for review. Flags appear inline in the transcript (struck through once applied) with the same accept/reject popover (over MCP: list_transcript_flags, review_transcript_flags, add_transcript_correction, add_transcript_cut). Captions drop accepted disfluencies, and they are always cut from the audio/video with jump cuts. Long pauses are shortened too (project setting Pauses: natural 0.6 s, tight 0.35 s, or keep), detected with Silero VAD on the clip audio. Cut boundaries are not taken from the chunk-level ASR timestamps (they drift by up to a word): every cut is first located by CTC forced alignment (wav2vec2, exact word boundaries at 20 ms from the audio plus the known words, ~0.2 s on CPU) and verified with one transcription of the seam; only if that fails is it re-aligned by transcribing windows around it (TRANSCRIPT_ALIGN_MODEL) twice with different offsets (word timings shift with the window), boundaries snap to the quietest 5 ms frame beside the neighbouring words, and a candidate search keeps the first seam that passes verification: both neighbours heard back to back, nothing left between them, and the deleted words gone (or, when the recogniser fills them back in from context, the removed audio on its own is heard as those words). Unverifiable cuts are skipped and listed on the clip. Seams get a 20 ms equal-power crossfade (acrossfade) with length-compensated trims so audio and video stay in sync. Substitution repairs between function words ("into between", "a the") are caught by a rule; legitimate compounds ("out of", "up to") are excluded.

Eval

Regression evals live in one runner with pluggable suites (selection, framing, asr-loops, scenes, pauses, alignment, speakers, overlays, cuts, transcription, disfluency, decisions, selection-model); fixtures are small and in git, large media is fetched into a gitignored cache, baselines are versioned and a run fails on regressions. See docs/EVALS.md, including how to turn a new bug into a case.

pnpm eval                 # all suites
pnpm eval cuts framing    # some suites
pnpm eval --offline       # skip suites that need models
pnpm eval --update        # accept new numbers as the baseline

The transcription suite synthesizes its cases with macOS text-to-speech so the ground truth is exact (eval/generate-audio.sh), and reports WER, entity recall and false entity insertions per model × vocabulary bias × LLM correction.

Results on the bundled cases (43 entity mentions, Speaches on CPU, qwen2.5:14b as judge):

Takeaway: vocabulary bias is the lever; the model size buys a little WER; the correction pass is the safety net for names the hints did not cover (and the review UI catches the rest).

Testing without AI keys

scripts/mock-ai.mjs is a tiny OpenAI-compatible server (chat completions + transcription with word timestamps). Point the app at it to exercise the whole pipeline offline:

OPENAI_BASE_URL=http://localhost:4545/v1 OPENAI_API_KEY=mock AI_MODEL=openai:mock TRANSCRIPTION_MODEL=openai:whisper-1

Troubleshooting

Start with docker compose exec app reelsaw status <projectId>; see docs/TROUBLESHOOTING.md for the CLI, structured logs, diagnostics and evals, including a "setting has no effect" checklist.

Roadmap

See docs/RESEARCH.md for the competitor audit and research review behind these priorities. Top items still open:

• Punch-in zooms on the emphasis words.
• Active speaker: more labelled two-way conversation cases; a learned audio-visual model (Light-ASD ONNX) if the synchrony heuristic plateaus. Saliency fallback when no face is visible.
• Direct publishing to YouTube Shorts / TikTok / Instagram.

License

Reelsaw is free software under the GNU Affero General Public License v3.0. You can run it, change it and host it; if you offer a modified version as a network service you must make your modifications available under the same licence. The hosted service's additions (billing) live in apps/server/src/ee and apps/web/src/ee under the Reelsaw Enterprise License and load only with REELSAW_EE=true; see LICENSING.md. Contributions need a DCO sign-off (git commit -s), see CONTRIBUTING.md.

Copyright (C) 2026 osbytes