MCP is the sole entry point. All invocations go through the MCP server (node mcp/server.mjs, or node mcp/client.mjs <tool> '<json>' from a shell). There is no web UI or HTTP API — the instagram-reel-generator.mjs engine refuses to run unless the MCP server spawns it (REEL_VIA_MCP). Start with VIDEO-DIRECTOR-SKILL.md (how to direct a reel) and FOOTAGE-MCP-SKILL.md / mcp/README.md (the tool catalog).

Spoooler turns raw footage, an Instagram URL, or a one-line topic into a finished, post-ready 1080×1920 reel — orchestrated tool-by-tool by whatever AI coding assistant you drive it with over MCP.

It can:

• expose that workflow over MCP so Codex and other MCP hosts can drive it tool-by-tool

• scrape real product media

• collect stock backgrounds

• generate voiceover with Kokoro or your own cloned pocket-tts/TADA voices

• align captions with whisper.cpp

• render the final MP4 with Remotion

What works out of the box vs. what's bring-your-own

None of these are required to try the tool — --skip-transcribe --transcript "..." gets you a full render with zero external transcriber or voice setup.

System requirements

• Node.js 20 or 22

• npm 10+

• Python 3.11+

• ffmpeg and ffprobe

• Git

Optional, only if you use these features:

• a Scrapling-capable Python venv (product/brand media scraping)

• pocket-tts on PATH (cloned-voice narration)

• an MLX-TADA setup on Apple Silicon (alternate cloned-voice engine)

• an MCP-compatible transcriber (Instagram URL / video upload input)

You'll also need API keys for whichever of these you want to use: Google Gemini, NVIDIA NIM, Groq, Pexels, Unsplash, logo.dev.

Quick start

git clone <this-repo-url>
cd instagram-reel-tool
npm install
cp .env.example .env

Fill in .env with the keys you want (see "Environment variables" below — most are optional and the tool degrades gracefully without them).

Verify the main packages resolve:

node -e "require.resolve('@modelcontextprotocol/sdk')"
node -e "require.resolve('remotion')"
node -e "require.resolve('@remotion/install-whisper-cpp')"
node -e "require.resolve('zod')"

Run the offline smoke test (no Instagram download, no LLM calls, no TTS):

node instagram-reel-generator.mjs \
  --skip-transcribe \
  --skip-tts \
  --offline \
  --transcript "Stop automating random tasks. The best AI systems start by finding the workflow bottleneck. Then they remove one handoff and measure the result." \
  --topic "AI workflow automation for founders"

This writes run artifacts under runs/<slug>/.

Environment variables

Copy .env.example to .env. Key variables:

• GOOGLE_API_KEY, GEMINI_MODEL — script generation (primary)

• NVIDIA_API_KEY, NVIDIA_MODEL — script generation (tried before Gemini)

• GROQ_API_KEY, GROQ_MODEL — script generation (final fallback)

• KOKORO_API_URL, KOKORO_API_KEY, KOKORO_MODEL, KOKORO_VOICE, KOKORO_SPEED — Kokoro TTS

• PEXELS_API_KEY, UNSPLASH_ACCESS_KEY — stock media

• LOGO_DEV_PUBLIC_KEY, LOGO_DEV_TOKEN — brand logo fetching

• IG_TRANSCRIBER_ROOT — optional, absolute path to your own MCP-compatible transcriber (see below)

• SCRAPLING_PYTHON — path to a Python interpreter with Scrapling installed

• FFMPEG_PATH, WHISPER_MODEL — rendering/captions

• POCKET_TTS_VOICE, POCKET_TTS_TONE, POCKET_TTS_QUALITY — cloned-voice narration

• TADA_* — alternate cloned-voice engine (see below)

None of these are required just to install and run the offline smoke test above.

Optional: Instagram URL / video transcription — pairs with ReelRecon

Transcribing an Instagram reel URL or an uploaded video file requires a separate MCP-compatible transcriber that this repo does not include. It must expose a run_mcp_server.sh script and a transcribe_input tool (input URL/path → transcript text).

ReelRecon — a companion tool, also by this author — is a drop-in fit: it transcribes Instagram profiles, direct video URLs, or uploaded audio/video with Whisper, and ships its own run_mcp_server.sh + transcribe_input MCP tool with exactly this interface. Typical pairing:

git clone https://github.com/4nw3rprod/ReelRecon.git

IG_TRANSCRIBER_ROOT=/absolute/path/to/ReelRecon

Feed a raw Instagram reel into ReelRecon for a clean transcript, then hand that transcript to Spoooler (via transcribe_source / --transcript) to script, voice, caption, and render the derivative reel — two focused tools instead of one that tries to do both.

Without it, Spoooler still works fully via --transcript or a plain --topic — you just skip the "give me a URL" step and provide the script input directly.

Optional: cloned-voice narration

The tool supports two cloned-voice backends. Neither ships with any voice data — bring your own.

Pocket-TTS (Kyutai)

1. Install the pocket-tts CLI and verify pocket-tts --help works.

2. Create a voice embedding (.safetensors) with whatever tooling you use to produce Kyutai-compatible embeddings.

3. Place it under audio/pocket-tts/voices/ (sibling to this repo, i.e. ../audio/pocket-tts/voices/) along with a voices.json index:

[
  {"id": "my-voice", "name": "My Voice", "embeddingFile": "my-voice.safetensors"}
]

4. Set POCKET_TTS_VOICE=audio/pocket-tts/voices/my-voice.safetensors in .env, or pass --voice-file per run.

If voices.json is missing or malformed, the MCP list_voices tool simply returns no cloned voices — Kokoro presets still work.

TADA (Hume MLX, Apple Silicon)

Optional alternate voice-cloning engine, clones from a short reference audio clip + its transcript instead of a pre-trained embedding.

python3 -m venv .venv-tada
source .venv-tada/bin/activate
python -m pip install --upgrade pip setuptools wheel
pip install mlx-tada

TADA_PYTHON=/absolute/path/to/.venv-tada/bin/python3
TADA_MODEL=HumeAI/mlx-tada-1b
TADA_PROMPT_AUDIO=/absolute/path/to/your/reference.wav
TADA_PROMPT_TEXT=
TADA_REFERENCE_CACHE=/absolute/path/to/instagram-reel-tool/.cache/tada/default-reference.npz

Notes:

• Apple Silicon only; follows Hume's apple/ implementation.

• Depends on the gated Meta Llama 3.2 base models on Hugging Face for the tokenizer.

• TADA_PROMPT_AUDIO (your own reference clip) is required; TADA_PROMPT_TEXT is optional.

• Use TADA_MODEL=HumeAI/mlx-tada-1b (English) or HumeAI/mlx-tada-3b (multilingual).

• Set TADA_WEIGHTS/TADA_TOKENIZER to force local weights/tokenizer instead of Hub downloads.

• Trigger via voiceEngine=tada in the MCP synthesize_voice tool, or --voice-engine tada on the CLI.

• Wrapper script: scripts/tada-tts.py.

Optional: product/brand media scraping (Scrapling)

The media-scraping pipeline uses Python and Scrapling.

python3 -m venv .venv-scrapling
source .venv-scrapling/bin/activate
python -m pip install --upgrade pip setuptools wheel
pip install "scrapling>=0.4,<0.5"
python -m playwright install

SCRAPLING_PYTHON=/absolute/path/to/instagram-reel-tool/.venv-scrapling/bin/python3

Verify:

source .venv-scrapling/bin/activate
python - <<'PY'
from scrapling.fetchers import Fetcher, DynamicFetcher, StealthyFetcher
print("scrapling fetchers ok")
PY

If Scrapling isn't set up, product scraping degrades to stock-media-only rather than failing the run.

MCP server

This repository includes an MCP server at mcp/server.mjs.

Local smoke tests

node mcp/test-client.mjs
node mcp/test-strategy.mjs

These confirm the server starts, tools register, and the strategy fast path works without an LLM script-generation step.

MCP host configuration

Most MCP hosts use a JSON mcpServers config, e.g. for a generic host settings file:

{
  "mcpServers": {
    "instagram-reel-tool": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/instagram-reel-tool/mcp/server.mjs"]
    }
  }
}

Codex configuration

Add to ~/.codex/config.toml:

[mcp.instagram-reel-tool]
command = "node"
args = ["/ABSOLUTE/PATH/TO/instagram-reel-tool/mcp/server.mjs"]

See mcp/README.md for the full tool catalog and more host examples.

Full render check

Once keys and (optionally) voices are configured:

node instagram-reel-generator.mjs \
  --skip-transcribe \
  --transcript "This is a short test reel about AI workflow automation." \
  --topic "AI workflow automation" \
  --render

Notes:

• the first whisper.cpp alignment run can take a while because the model installs into .cache/whisper-align

• the first cloned-voice run can take longer while pocket-tts loads

• if no media APIs are configured, renders complete but with empty or degraded media layers

Troubleshooting

list_voices returns no cloned voices

Check ../audio/pocket-tts/voices/voices.json and the referenced .safetensors files exist.

Scraping returns nothing

Check SCRAPLING_PYTHON points to a working venv, Scrapling imports successfully, Playwright's browsers are installed, and you have network access.

Video upload fails to transcribe

Check IG_TRANSCRIBER_ROOT points to a working transcriber exposing run_mcp_server.sh + transcribe_input, ffmpeg works, and the uploaded file is under 200 MB.

Cloned voice generation fails

Check pocket-tts is on PATH, the selected .safetensors file exists, and voices.json references the correct embedding filename.

Captions do not align

Check ffmpeg is installed, npm install completed successfully, @remotion/install-whisper-cpp resolves, and the first whisper model install was allowed to complete.

Render succeeds but visuals are empty

Check PEXELS_API_KEY/UNSPLASH_ACCESS_KEY are set and stock/scraped media actually downloaded into the run folder.

Files worth reading

• README-instagram-reel-generator.md

• mcp/README.md

• SKILL.md

• instagram-reel-generator.mjs

Related project

• ReelRecon — Instagram/video transcription over MCP or web UI. See Optional: Instagram URL / video transcription above for how the two fit together.

Roadmap: a closed-loop, autonomous content pipeline

Spoooler and ReelRecon already cover transcribe → script → produce end to end over MCP. The next integration closes the loop from raw footage to a published post, with no manual handoff in between:

The goal: point the pipeline at a source once, and let it run end to end — discovery through publishing — as a single MCP-orchestrated workflow. No manual export/upload step, no juggling separate scheduling tools, no re-planning content by hand. Set it up once, then stop thinking about posting and go back to thinking about what's worth making.

This is planned, not yet wired up — tracking here so the shape of the integration is public before the code is.

Built with

Spoooler is a thin orchestration layer over a handful of open-source projects doing the real work:

Full dependency list in package.json.

License

MIT — see LICENSE.