Skip to main content
Glama
harshil-1411

engine

by harshil-1411
OverviewSchemaSourceRelated ServersDiscussionsAdmin
TypeScript
Local

https://github.com/user-attachments/assets/d860c1b2-8653-455c-9bfa-640ad8a0b849

Turn a README, a paper, a web page or a folder of clips into a finished, captioned short video, one package per platform, without leaving Claude Code.

Features · See what it makes · Quick start · Requirements · How it works · Commands · Limits · What's new

Why

Turning knowledge into short videos usually means a timeline editor, a caption tool, a thumbnail tool and a checklist for every platform's safe zones and limits. video-studio compiles instead:

  • Grounded: every claim on screen cites a line in your sources. verify shows what is covered.

  • Platform-ready: one dist/<platform>/ package each for TikTok, Instagram Reels, YouTube Shorts, LinkedIn and Facebook, with the video, cover, captions, post copy and a QA report. lint checks captions and text against each app's UI.

  • Made to be watched: plans follow a story arc with timed "reads", each list item, step or number appears as the voice says it, and Claude reviews contact sheets of its own render before handing it over.

  • Works with real footage: ingest recordings or YouTube/Vimeo/Loom links, transcribe them locally in ~99 languages, find the best short clips, keep a moving speaker in frame for vertical video, cut away to graphics while they talk, and trim pauses and filler words.

  • You stay in control: Claude proposes and waits for your approval; paid voices, model downloads and screen recordings ask you through Claude Code's approval dialog and follow your policy.yaml spend limits; renders can be cancelled.

  • Reproducible: video.lock pins every tool, font, renderer and asset hash. Re-renders are cached scene by scene, and diff and test catch regressions.

  • Local-first: ffmpeg, system text-to-speech and local whisper.cpp. Claude writes the plan, so the plugin needs no LLM API key, and none of the features below need a paid service.

Related MCP server: MotionKit MCP Server

Features

Highlights

  • 🎬 Motion written as code. Claude can write a scene as an HTML page drawn by a pure seek(t) function: springs, morphs, match cuts and kinetic type at the level of hand-made motion design. Every page runs under a strict security policy, is checked for unsafe code before it renders, and is proven deterministic (the same time always draws the same frame).

  • 📏 Pacing you can measure. QA counts big visual changes per second, the longest still stretch and frozen time, and fails a slideshow-paced reel. Vague asks like "make it pop" become acceptance numbers the render must meet, and compare scores your render against a reference video you like.

  • 🎯 Style from a reference reel. analyze measures how a reel you like moves (entrance times, easing, stagger, holds) without keeping anything from it, and write_style turns that into a style pack in your project.

  • 🚀 Launch videos in one command. /video-studio:launch turns your repo or site into an 18–22 s reel of the product in use, in its own colours, fonts and logo (drafted from the source for you to accept), with a synthesized score, sound effects and a poster on frame 0 for chat previews. One approval, after the preview.

  • 🎵 Music that drives the cut. Beat analysis finds beats, downbeats and the drop; cuts snap to them and sound effects land on their peak. No track? synth: scores are composed locally, CC0, with an exact beat grid.

  • 📱 One source, every platform. Per-platform packages (video, cover, captions, post copy, QA) for Instagram Reels, YouTube Shorts, TikTok, LinkedIn and Facebook, with text and captions kept clear of each app's UI.

  • ✅ Grounded and reproducible. Every on-screen claim cites your sources (verify), video.lock pins every tool and asset, and scenes re-render only when something they use changes.

  • 🔒 Local-first, no keys needed. ffmpeg, system voices and local whisper. Claude writes the plan, so there is no LLM API key; paid providers are optional placeholders until you add keys.

Everything it does

Area

Features

Sources

Markdown, text, PDF, DOCX, PPTX, web pages (JavaScript-built pages too, with your approval), local repos, video/audio files, clip folders, video URLs (YouTube, Vimeo, Loom via your yt-dlp)

Planning

Story-arc plans with hooks and a hook-strength check, 24 templates that ask for the inputs they need first, a beat-level plan at the approval step, series bibles for recurring characters and looks, 7 tone presets (polished, playful, deadpan, cinematic, energetic, app-store, parody), product videos planned around the product in use, a brand kit drafted from your repo or site (brand_draft), A/B variants (hooks, covers, 15 s / 30 s cuts)

Visuals

16 scene kinds including Claude-written motion pages, 4 style packs with banned-effect lists plus project styles measured from a reference reel, brand kits with your project's own fonts, camera moves, transitions, word cues that land graphics on spoken words, text reveals timed to the beat and held long enough to read, count-ups, optional motion blur, a poster baked into frame 0

Review before render

stills sheets at chosen times, beats or downbeats (with the music and reveal timing the render uses); a determinism and loop-seam check for every motion page; review tiles at every transition

Audio

System TTS or ElevenLabs, 4 CC0 beds plus locally synthesized scores, beat and downbeat snapping, a synthesized CC0 sound-effect library, sound effects on their peak, motion that reacts to the music, ducking, −14 LUFS

Footage

Local transcription (~99 languages, speaker turns) with a glossary for names, best-clip shorts, subject tracking for vertical reframes, cutaways, tighten for pauses and filler words (paced like your own edits, every join checked for clipped words), redaction, letterbox removal, an audio offset for recordings whose sound lags the picture

Captions and languages

Phrase captions clear of platform UI, keyword emphasis, sound-event captions, localize with Devanagari, Japanese and Arabic fonts, RTL and CJK line breaking

Checks

30+ lint rules (UI zones, contrast, reading speed, caption sync, insert timing, cuts on the beat, story arc, title length, brand rules, stock phrases, busy crossfades, reveals too fast to read, sound-effect licence and placement, banned effects, acceptance numbers, loop seams, unsafe motion pages), technical QA (loudness, black, frozen, motion density, share of frames moving, loop seam, flashing, A/V sync), review contact sheets, compare against a reference

Export

Per-platform packages, C2PA signing, and an editable timeline (import-tested) for DaVinci Resolve or Final Cut (FCPXML and OTIO)

Trust and control

Claim verify, video.lock, golden-frame test, diff, provenance, optional C2PA signing, policy.yaml spend limits and consent, render_cancel

Generative (prep)

Shot cards compiled into ready-to-paste prompt packs for Seedance, Veo, Kling, Wan, Runway and Hailuo, offline with no spend; provider and publishing keys are optional placeholders until Phase 7/9

See what it makes

Everything below was made by the plugin itself: no hand editing.

This README as a reel. docs/media/hero.mp4 is a narrated 1080×1920 final render made from this README. It uses the animated-explainer structure, the technical style and a macOS Premium voice, and every line cites a README line. The project that makes it is examples/readme-hero:

The strips below are frames from preview renders (built-in ffmpeg renderer) made during development.

Built-in scene kinds. The strip shows kinetic text, a stat, a timeline, before/after, a quote, a map and a lower third, from a text-over-music reel with no voiceover:

4 style packs, on the same content. From left: the default, minimal, editorial, technical and energetic:

Languages. The Whisper paper (arXiv 2212.04356) as Hindi (top) and Japanese (bottom) versions made with localize: shaped Devanagari, Japanese line breaking and script-aware captions and covers:

Your own footage. A folder of clips cut to the beat of a bundled music bed, with text over the footage (aesthetic-broll and silent-vlog):

Quick start

Requirements: Claude Code, Node.js 22.13+ and FFmpeg (brew install ffmpeg on macOS). Nothing else is required: the engine is a single bundled file. See Requirements for the full list, and check your machine with /video-studio:doctor.

Inside Claude Code:

/plugin marketplace add harshil-1411/claude_plugin_video_studio
/plugin install video-studio@video-studio-marketplace

Then:

/video-studio:create README.md as a 30-second 9:16 reel for instagram and youtube-shorts

Claude reads the source, proposes a hook, a scene plan and a storyboard, and waits for your approval. It then renders a preview, looks at it, then the final, and writes the packages:

dist/
├── reel.mp4  clean-master.mp4  captions.srt  captions.vtt  cover.jpg
├── video.lock  render-manifest.json  provenance.json  video-spec.json
├── instagram/        video.mp4  cover.jpg  captions.*  post.json  qa.json
└── youtube-shorts/   …

  • Voice: macOS picks your best installed voice (add a Premium voice in System Settings → Accessibility → Spoken Content). An ElevenLabs key (/plugin → video-studio → Configure, stored in the OS credential store) is used only when your policy.yaml allows it or you ask for it.

  • HyperFrames renderer (needed for motion scenes): install Google Chrome, then run /video-studio:doctor: it prints the one-time install command (the pinned @hyperframes/producer, installed into the plugin's data folder) and confirms Chrome starts. Nothing is installed until you run it. Without HyperFrames, every scene kind still renders with ffmpeg, but a motion scene appears as a labelled text stand-in and the render says so.

  • Video URLs: brew install yt-dlp to ingest YouTube, Vimeo or Loom videos (subtitles become the transcript).

  • Transcription: whisper.cpp (brew install whisper-cpp); models are downloaded only after you approve.

Nothing needs a key. /plugin → video-studio → Configure shows these fields; secrets are kept in the OS credential store and only reach the engine, never a shell:

Kind

Keys

In use today

ElevenLabs (voiceover), used only when your policy.yaml allows it or you ask for it

Placeholders for Phase 7 (AI video)

Runway, HeyGen, fal.ai, Kling, Google Gemini (Veo), BytePlus ModelArk (Seedance), Alibaba DashScope (Wan), MiniMax (Hailuo)

Placeholders for Phase 9 (publishing)

YouTube client ID and secret, Meta (Instagram) token, LinkedIn token, TikTok client key and secret

Placeholders can be filled in any time; they have no effect until their integration ships. /video-studio:doctor shows which keys are set, never their values.

  1. Give Claude the reference video with your source: /video-studio:create my-notes.md as a 20-second 9:16 reel that feels like reference.mp4. The plan skill asks for a reference, a photo and your brand first, and turns "make it feel like this" into acceptance numbers (big changes per second, frozen %, holds).

  2. /video-studio:analyze reference.mp4 write_style ref-look measures the reference's motion timing (entrance length, easing, stagger, holds) and saves it as a style pack in your project. Only timing is kept, never its words, frames or audio.

  3. After the preview render, /video-studio:stills shows each scene on the beats before the final render, and /video-studio:compare scores your render against the reference (frozen %, changes per second, cut rate, loudness).

  4. QA fails a render that misses the acceptance numbers, so a slideshow-paced reel cannot pass silently.

git clone https://github.com/harshil-1411/claude_plugin_video_studio.git video-studio
cd video-studio && pnpm install
claude --plugin-dir .

Requirements

/video-studio:doctor checks all of this on your machine and says exactly what is missing and how to fix it.

Operating system

OS

Status

macOS (Apple silicon or Intel)

Fully supported and tested. Everything works, including macOS voices and automatic subject tracking (Apple Vision).

Linux

Supported. Voice uses espeak-ng; subject tracking is done by Claude from shot sheets instead of automatically.

Windows

Not tested. There is no built-in system voice (use no voice or ElevenLabs), and Chrome is not found automatically (set CHROME_PATH). WSL2 with Linux is the safer route.

Required software

Tool

Why

Install (macOS)

Claude Code

Hosts the plugin; Claude writes the plans and pages

see the Claude Code docs

Node.js 22.13+

Runs the engine (it uses Node's built-in SQLite)

brew install node

FFmpeg and ffprobe, with libx264 and libass

Every render, caption burn-in and QA check. libass (with fribidi) is needed for Arabic, Hebrew and Devanagari captions

brew install ffmpeg

Optional software (each unlocks one feature; nothing is installed for you)

Tool

Unlocks

Google Chrome + the HyperFrames producer

motion scenes (Claude-written code), richer motion graphics, stills, the determinism check, and opening JavaScript-built pages at ingest (render_js). doctor prints the one-time install command

whisper.cpp (brew install whisper-cpp) + a model (about 148 MB; 488 MB for speaker turns, downloaded only with your approval)

Transcribing your footage, shorts, tighten, and exact word timings for system voices

yt-dlp (brew install yt-dlp)

Ingesting YouTube, Vimeo and Loom links

A macOS Premium or Enhanced voice (System Settings → Accessibility → Spoken Content)

Natural-sounding narration with the free system voice

An ElevenLabs API key

Premium voiceover (optional and policy-gated; see API keys below)

FFmpeg with zscale (libzimg)

Tone-mapping HDR phone footage (without it HDR passes through with a warning)

Hardware

Minimum

Recommended

CPU

Any 64-bit CPU

4+ cores (scenes render in parallel, up to 2 at once)

Memory

8 GB

16 GB. Each parallel scene render wants about 1.5 GB free, and HyperFrames runs Chrome

Disk

About 25 MB for the plugin

1–2 GB free for your projects, caches and optional whisper models. A 20 s 1080p reel project is about 50 MB, and the shared cache grows by a few hundred MB over many projects

GPU

Not needed

Not needed: everything renders on the CPU

Internet

Only to install the plugin

Only for URL ingest, optional downloads, and paid providers you enable

How it works

flowchart LR
  A["Sources<br/>md · pdf · docx · pptx · url · repo<br/>video · audio · clip folders"] --> B["ingest<br/>ContentIR + evidence refs"]
  B --> C["plan<br/>brief · grounded VideoSpec · storyboard"]
  C -->|your approval| D["render<br/>voice · scenes · captions · music"]
  D --> E["lint + QA<br/>platform contracts · WCAG · loudness"]
  E --> F["dist/&lt;platform&gt;/<br/>video · cover · captions · post · qa"]

  • Claude is the creative engine. Skills guide Claude to write the brief and the scene spec.

  • The engine does the rest. A bundled MCP server validates, renders, runs QA and packages. It is deterministic, cached and has no LLM calls.

  • Three contracts connect the stages:

    • ContentIR: the sources, with provenance.

    • VideoSpec: a provider-neutral scene graph.

    • RenderManifest: exactly what happened.

    Their JSON Schemas are in schemas/.

  • Facts are data, not code: platform-specs/*.yaml record each app's limits and UI masks, and provider-specs/*.yaml each AI video model family's limits and prompt syntax, with source URLs and the date they were checked.

What you can make

Inputs

Markdown, text, PDF, DOCX, PPTX, web pages (a page built by JavaScript can be opened once in an isolated headless Chrome with render_js, after you approve it), local repos, video and audio files (local transcription in English or ~99 languages with detection, optional speaker turns for English conversations), folders of clips, video URLs (YouTube, Vimeo, Loom through your own optional yt-dlp, using their subtitles when present; direct .mp4/.mp3 links need nothing extra)

Templates (24)

explain · educational · listicle · faceless-listicle · product-launch · devtool-launch · product-demo · product-ui · case-study · before-after · carousel-story · animated-explainer · text-over-music · talking-head · aesthetic-broll · silent-vlog · oddly-satisfying · ambient-slice-of-life · ui-morph-loop · kinetic-type · ambient-loop · slides-narrated · topic-explainer-9 · product-hero; templates can ask for the inputs they need first (reference video, photo, real UI states)

Scene kinds (16)

typography · code · chart · stat · diagram · timeline · comparison · split_screen · quote · kinetic_text · lower_third · map · screenshot · cta · end_card, plus motion: a page Claude writes as code (seek(t), springs, the brand tokens and the music's beat grid), checked for safety and determinism before it renders, with optional motion blur; plus real footage with text overlays

Voice

macOS say (automatically picks an installed Premium/Enhanced voice; with local whisper installed its word timings are aligned to the audio, so captions and cues land exactly) or espeak-ng, ElevenLabs (optional key), no voice (text over music), or the speech already in your footage; pace set with voice.rate_wpm (default 160)

Audio

4 bundled CC0 music beds (ducked under speech), locally synthesized scores (synth:pulse, lofi, ambient, drive: CC0, exact beat grid), cuts snapped to beats or downbeats, native clip sound (with av_offset_ms when it lags the picture), crossfades, 18 synthesized CC0 sound effects (bundled:whoosh-soft, pop, click, chime…, each measured for brightness and level) placed on their peak, motion pages that can breathe with the music's energy, −14 LUFS with true-peak headroom

Footage

Crop, contain or blurred-pad fits, trim and speed, text overlays, automatic removal of baked-in letterbox bars, and redact regions to blur inboxes, names or dashboards in screen recordings; subject tracking that keeps a moving speaker in frame when a landscape video becomes vertical (footage_focus, macOS Vision); footage_look shot sheets so Claude sees the footage before choosing clips; cutaways from a talking head to a graphic while the speaker keeps talking; a brand or series glossary that fixes misheard names in transcripts and captions; tighten paced from a video you edited, with every cut checked for clipped or repeated words; quality warnings (dark or bright picture, clipped or unclear audio); rotated phone video and HDR handled

Captions

3–7 word phrases on plates, placed clear of each platform's UI, held long enough to read, broken at speaker changes, with keyword emphasis and sound-event cues like [music]; turn them off per scene where kinetic text already shows the words

Looks

Style packs (minimal, editorial, technical, energetic, or your own in <project>/styles/, e.g. measured from a reference reel by analyze write_style) and brand kits (colours, fonts, weights, motion, a corner logo, forbidden treatments, drafted from your repo or site by brand_draft with evidence for each value, fonts from your project's fonts/ folder, banned phrases, pronunciation overrides such as LLM → "L L M" that keep captions as written); scene transitions (crossfade, fade to black, slide, zoom, whip) that keep narration in sync; per-scene camera moves (push in, pull out, punch, reveal, drift, hold); word cues that land each list item, step or number on the word that says it

Languages

localize translation sheets; bundled Noto fonts for Japanese, Devanagari and Arabic; CJK line breaking; right-to-left text

Checks

30+ lint rules (platform UI zones, contrast, reading speed, caption timing, cues, data inserts on the words that say them, story arc, cutaway rhythm, title length (a heuristic), brand rules, banned effects, acceptance numbers, loop seams, footage quality, stock phrases, busy crossfades, reveal speed, sound effects), technical QA (loudness, black or frozen frames, motion density, share of frames moving, flashing (approximates WCAG 2.3.1; red flashes not measured), A/V sync), stills sheets before a render, review contact sheets with problem scenes bordered, an automatic review → fix → re-render loop, and compare before/after pages or against a reference video with metrics

Trust

verify claim coverage, video.lock, golden-frame test, diff, provenance, optional C2PA content credentials (export sign); secrets found in sources are redacted

Control

policy.yaml (allowed providers, spend limits, approval threshold), consent recorded in project/consent.json, render_cancel, one render per project at a time

Examples

Example

What it shows

examples/readme-hero

The narrated hero reel above, grounded in a README snapshot

examples/text-to-motion-graphic

A 30 s explainer from Markdown notes (also the golden-frame test)

examples/reel-grammar

Every Phase 5 scene kind, the energetic style and a music bed, with no voiceover

examples/code-motion-loop

A 6 s seamless UI-morph loop written as a motion page on a synthesized score, cut to downbeats (needs HyperFrames)

examples/launch-video

A 20 s launch reel for an invented product, made the /video-studio:launch way: brand drafted from its site, the product flow as five motion pages, a synthesized score, sound effects and a poster on frame 0 (needs HyperFrames)

examples/demo-app

A tiny web app to try /video-studio:demo screen recording on

Commands

Command

What it does

/video-studio:create

The whole flow, from a source or an idea to packages, with an approval step

/video-studio:launch

A short launch reel of something you built, from its repo or URL: the product in use, its own brand, one approval after the preview

/video-studio:plan · validate

Brief, grounded spec and storyboard, built on a story arc (hook, open loop, escalation, payoff, CTA); explains every validation issue

/video-studio:render · qa · export

Local render (preview, then final; cancel anytime), technical QA (including flashing and A/V sync), per-platform packages (sign for C2PA; timeline for a DaVinci Resolve or Final Cut project)

/video-studio:lint · verify

Platform contract checks with a fix loop (UI zones, caption readability and sync, cuts on the beat, story arc); claim coverage against the sources

/video-studio:stills

Frames of each scene at chosen times, beats or downbeats, before the full render

/video-studio:prompt-pack

Prompts for Seedance, Veo, Kling, Wan, Runway and Hailuo compiled from shot cards (offline: nothing generated or spent)

/video-studio:review · compare

Contact sheets, frame strips, transition tiles and crops of a render (lint findings bordered), so Claude looks at the video before handing it over; a before/after page that plays two versions in sync (side by side, stacked or wipe)

/video-studio:test · diff

Golden-frame regression tests; spec, lock and frame diffs between renders

/video-studio:variants · adapt

Hook × cover A/B sets with an experiment manifest; new aspect, length or platform

/video-studio:localize

Language versions from a translation sheet, re-timed for the language

/video-studio:ingest · shorts · analyze

Documents, web pages (JavaScript-built ones with your approval), repos, media files and video URLs; local transcription (language detection, speaker turns); standalone clips from a long talk, with shot sheets, subject tracking and cutaways; a reference video's format, motion timing and speech pacing (write_style saves the timing as a project style)

/video-studio:tighten

Cleans up talking-head footage: shortens pauses (optionally paced like a video you edited), cuts filler words and drops retakes, and checks every join for clipped words (dry run first, new asset on apply)

/video-studio:demo

Records a scripted walk through your running app (inputs are blurred)

/video-studio:doctor

Checks ffmpeg, fonts, Chrome, whisper, HyperFrames and keys

What it does not do (yet)

  • No generative video or avatars yet. Adapters are planned (Phase 7, needs keys; the key fields already exist in Configure and do nothing until then). Until then those scenes render as titled placeholder cards, and prompt_pack writes ready-to-paste prompts for each generator. Sora is intentionally not supported.

  • No posting or analytics. It produces packages and post copy; you upload them. Platform "trending sounds" are added in each app, and post.json reminds you of that.

  • HyperFrames is optional, except for motion scenes. The built-in ffmpeg renderer covers every other scene kind. motion pages (Claude-written code) need HyperFrames and Google Chrome; without them they render as a reported text stand-in.

  • Some checks are approximations. Flash detection measures average brightness (it follows WCAG 2.3.1 but does not measure red flashes); the title-length band is a rule of thumb, reported as a warning only; motion density counts sudden changes, and moving_pct counts frames that move at all (smooth motion and crossfades included), so a very slow drift can sit near its threshold.

  • The editor timeline is a starting point. It places every scene on its exact frame and keeps the audio and captions, but transitions become markers rather than rebuilt dissolves.

  • Whisper models are downloaded only with your consent (about 148 MB; 488 MB for the speaker-turn model). You can supply SRT/VTT captions instead.

  • Speaker turns are English-only and label two alternating speakers (S1/S2); rename them if there are more.

  • Subject tracking is automatic on macOS only. Elsewhere Claude marks the subject from shot sheets.

  • No colour grading yet. Dark or bright footage gets a warning, not a fix.

  • Demo capture never starts your app. You start it and give the URL, and every step is approved first.

  • SVG logos can't be the corner logo (ffmpeg can't read SVG); use a PNG. Project fonts must be TTF or OTF.

The roadmap is in docs/PLAN.md and the current state in docs/HANDOFF.md.

pnpm install
pnpm typecheck        # tsc -b
pnpm test             # vitest
pnpm schemas          # regenerate schemas/*.schema.json
pnpm bundle           # build dist/mcp.mjs (single-file ESM, committed)
pnpm smoke            # start dist/mcp.mjs over stdio and check its tools
claude plugin validate --strict .claude-plugin/plugin.json   # plugin + skills
claude plugin validate --strict .                            # marketplace
pnpm check            # all of the above plus golden frames, stopping at the first failure
pnpm hooks            # once: run pnpm check before every git push

After changing anything under packages/, rerun pnpm bundle and commit dist/mcp.mjs (pnpm check fails if the committed bundle is stale).

Package

Role

packages/schema

zod models → schemas/*.schema.json

packages/core

project folders, cache, SQLite ledger, jobs

packages/ingestion

extractors (documents, web, repos, media)

packages/media

ffmpeg, audio mix, captions, QA, ASR, beat detection

packages/renderer

ffmpeg, footage and HyperFrames renderers; tokens, styles, scripts

packages/platforms

platform contracts and layout zones

packages/voice

TTS backends

packages/prompts

provider specs and prompt compilers for shot cards

packages/mcp

the MCP server (dist/mcp.mjs) and every tool

Data lives next to the code: skills/, templates/, styles/ (a project can add its own in <project>/styles/), music/, sfx/, fonts/, platform-specs/, provider-specs/ and research-specs/.

Tests that need real Chrome or a whisper model are skipped by default:

VS_TEST_RENDER=1 npx vitest run packages/renderer packages/mcp/src/stills.test.ts   # real HyperFrames/Chrome renders
VS_TEST_RENDER=1 VS_UPDATE_GOLDEN=1 npx vitest run tests/golden-frames              # record the motion example's goldens
VS_TEST_RENDER=1 npx vitest run packages/mcp/src/render-page.test.ts                # render_js isolation on a local JS page
VS_TEST_WHISPER_MODEL=<path to a ggml model> npx vitest run packages/mcp/src/tighten.test.ts
VS_DEBUG_CAPTURE=1 …                                                               # trace every Chrome capture step

Contributing

Guides for adding archetypes, style packs, platform packs and providers are in docs/contributing/. Ingested content is always treated as untrusted data: the plugin never executes code from sources in its own process. The one exception is opt-in render_js, which opens a page once in an isolated headless Chrome after you approve it, with every request checked by the engine and nothing clicked.

Credits

Some ideas re-expressed from latent-spaces/brag (MIT).

License

Apache-2.0. See LICENSE. The bundled fonts are OFL-1.1 (see fonts/README.md), the bundled music beds are CC0 (see music/README.md), and the bundled sound effects are CC0, synthesized by the plugin (see sfx/README.md).

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Resources

Related MCP Connectors

Related MCP Servers