@samuraizer/cli

@samuraizer/cli

Local-first CLI that turns meeting recordings into transcripts, summaries, action items, and decisions — entirely on your machine. No cloud, no subscriptions, no data leaving your network.

Reference implementation of the memnex specification. All outputs conform to memnex v0.2, including a full provenance chain.

💻 System Requirements

RAM	Recommended model
8 GB	qwen2.5:3b
16 GB	qwen2.5:7b
32 GB+	qwen2.5:14b (default)

Apple Silicon (M1/M2/M3/M4) and recent x86 CPUs with AVX2 are recommended. Whisper transcription is CPU/Metal-accelerated; LLM inference uses Ollama's defaults.

⚙️ Prerequisites

Install the required tools:

• Node.js ≥ 20 — nodejs.org
• ffmpeg — for audio processing
• whisper-cli — from whisper.cpp
• Ollama — ollama.com

Start Ollama and pull a model:

ollama serve
ollama pull qwen2.5:14b

📦 Installation

npm install -g @samuraizer/cli

Migrating from the legacy samuraizer package? Versions ≤ 0.2.0 of the unscoped samuraizer package on npm are deprecated. Run npm uninstall -g samuraizer && npm install -g @samuraizer/cli to migrate. The CLI binary on your PATH is still called samuraizer.

🚀 Quick Start

samuraizer init
samuraizer process meeting.m4a

On a 30-minute recording this typically takes 3–5 minutes on Apple Silicon and 8–15 minutes on x86 CPUs, depending on the model.

🧪 Commands

Process an audio file

samuraizer process meeting.m4a              # full pipeline
samuraizer process meeting.m4a --verbose    # show detailed metadata
samuraizer process meeting.m4a --force      # recompute all steps
samuraizer process meeting.m4a --verbose --force

Run individual steps

samuraizer normalize input.m4a output.wav   # normalize audio for Whisper
samuraizer summarize transcript.txt         # generate summary from transcript
samuraizer actions transcript.txt           # extract action items
samuraizer decisions transcript.txt         # extract decisions

Configuration

samuraizer init           # create default config file
samuraizer config path    # show config file location
samuraizer config get     # print resolved config as JSON

Other

samuraizer --help
samuraizer --version

⚙️ Configuration

Samuraizer uses a global JSON config file.

Config location

• macOS: ~/Library/Application Support/samuraizer/config.json
• Linux: ~/.config/samuraizer/config.json
• Windows: %AppData%/samuraizer/config.json

Example config

{
  "model": "qwen3.5:14b",
  "ollamaBaseUrl": "http://127.0.0.1:11434",
  "whisperCommand": "whisper-cli",
  "whisperModelPath": "/absolute/path/to/ggml-model.bin",
  "language": "en",
  "ffmpegCommand": "ffmpeg",
  "ffprobeCommand": "ffprobe"
}

Config fields

• model — LLM model used for analysis (summary, action items, decisions)
• ollamaBaseUrl — URL where Ollama is running
• whisperCommand — Command used to run Whisper
• whisperDevice (optional) — GPU/device whisper-cli runs on. Accepts a device index (0, 1), a comma-separated list ("0,1"), or a GPU UUID; value semantics match CUDA_VISIBLE_DEVICES. Omit to use the default device.
• ffmpegCommand — Command used for audio processing
• ffprobeCommand — Command used for audio inspection

Every config field can also be overridden with an environment variable, e.g. SAMURAIZER_WHISPER_DEVICE=1 samuraizer process meeting.m4a.

Selecting a GPU

If your machine has multiple GPUs, pin whisper transcription to a specific one:

{
  "whisperDevice": 1
}

Or per-run, without editing the config:

SAMURAIZER_WHISPER_DEVICE=1 samuraizer process meeting.m4a

📂 Example output

After processing, you'll find structured files in output/<recording-name>/:

output/meeting/
  transcript.txt
  summary.txt
  action-items.json
  decisions.json
  report.txt
  meeting.json

The meeting.json file is a memnex v0.2-conforming document combining all outputs with a full provenance chain.

summary.txt

Team standup focused on Q2 roadmap and infrastructure migration.
The frontend team will start the Next.js upgrade next week...

action-items.json

[
  {
    "owner": "Alice",
    "task": "Set up staging environment for migration testing",
    "deadline": "by end of week"
  },
  {
    "owner": "Bob",
    "task": "Review the auth refactor PR",
    "deadline": null
  }
]

decisions.json

[
  {
    "decision": "Adopt Next.js 15 for the new dashboard",
    "rationale": "Better SSR and built-in App Router support"
  }
]

🔁 Resume behavior

Samuraizer skips steps whose output files already exist. If processing crashes or you stop it mid-pipeline, just re-run the same command — completed steps are reused.

Use --force to recompute everything from scratch.

⚠️ Troubleshooting

Ollama not running

ollama serve

Ollama on a non-default port

Update ollamaBaseUrl in your config:

{
  "ollamaBaseUrl": "http://127.0.0.1:11500"
}

Out of memory during analysis

Switch to a smaller model:

ollama pull qwen3.5:14b

Then update model in your config to qwen2.5:7b (or qwen2.5:3b on machines with 8 GB RAM).

Model not found

Make sure the model in your config is actually pulled:

ollama list
ollama pull <model-name>

whisper-cli not in PATH

Build whisper.cpp and ensure the binary is on your PATH, or set the absolute path in whisperCommand in your config.

ffmpeg not found

macOS:

brew install ffmpeg

Linux:

# Debian / Ubuntu
sudo apt install ffmpeg

# Arch / CachyOS
sudo pacman -S ffmpeg

# Fedora
sudo dnf install ffmpeg

Windows:

winget install Gyan.FFmpeg

🤖 AI agent access (MCP)

Samuraizer also ships with a companion MCP server, @samuraizer/mcp-server, that lets AI agents (Claude Desktop, Claude Code, MCP Inspector) query your processed meetings and run the pipeline on demand.

📝 Changelog

See CHANGELOG.md for release history.

📄 License

MIT — see LICENSE.

🔗 Source code

Part of the Samuraizer monorepo: github.com/UladzKha/samuraizer-cli.