Crispasr Agent Transcriber MCP Server

crispasr-agent-transcriber

Local-only transcription for Codex and MCP-based AI agents, powered by CrispASR. No cloud uploads, no API keys required for transcription.

GitHub Release | npm installer | PyPI package | MCP Registry

What it does

Give it a local audio or video file. It:

1. Probes the spoken language (English or Chinese) using CrispASR's FireRed LID.
2. Starts a local CrispASR server with the right backend -- Cohere Transcribe for English, Qwen3-ASR for Chinese.
3. Extracts audio from video with ffmpeg when needed.
4. Calls CrispASR's /v1/audio/transcriptions endpoint.
5. Writes the transcript and metadata to disk.

Everything runs on your machine. Media never leaves it.

Quick install for Codex

The plugin includes the Codex Skill, command-line tool, and MCP server. Media stays on your computer. Model files are never downloaded automatically.

1. Install prerequisites

Install Node.js 20 or newer, uv, and ffmpeg. The installer uses uv to provide Python.

node --version
uv --version
ffmpeg -version

2. Run the installer

npx @emiyakatuz/crispasr-agent-transcriber@latest install

The installer:

• downloads the matching GitHub Release and verifies its SHA-256 checksum;
• installs the plugin under ~/plugins/crispasr-agent-transcriber;
• installs the Python and MCP dependencies;
• detects CUDA, Vulkan, or CPU and installs the best CrispASR build;
• registers the plugin in the Codex Personal marketplace;
• preserves existing models, binaries, and outputs during updates.

3. Add the local models

When a model is missing, the installer prints its official source and stops. Download the three files listed under Required models into:

~/plugins/crispasr-agent-transcriber/models/

Then verify the complete installation:

npx @emiyakatuz/crispasr-agent-transcriber@latest doctor

4. Enable the plugin

With a Codex build that supports plugin commands, run:

codex plugin add crispasr-agent-transcriber@personal

If the CLI has no codex plugin command, open the Codex desktop Plugins view and install CrispASR Transcriber from the Personal marketplace. Start a new conversation, then ask:

Transcribe C:\path\to\sample.mp4 with CrispASR using auto language detection.
Save a verbose JSON transcript and an SRT subtitle file.

Update or uninstall

npx @emiyakatuz/crispasr-agent-transcriber@latest update
npx @emiyakatuz/crispasr-agent-transcriber@latest uninstall

Uninstall preserves local models, CrispASR binaries, and outputs. Use uninstall --purge-data only when those files should also be deleted. See Plugin installation for manual installation and troubleshooting.

Direct command-line use

After installation, you can run the transcription script without Codex:

Set-Location (Join-Path $HOME "plugins\crispasr-agent-transcriber")
uv run python scripts/transcribe.py sample.mp4 --profile auto `
  --manage-server `
  --english-model models\cohere-transcribe.gguf `
  --chinese-model models\qwen3-asr-1.7b-q4_k.gguf `
  --lid-backend firered --lid-model models\firered-lid-q2_k.gguf `
  --format verbose_json

Use with other AI agents

The MCP server is the cross-agent interface. Any agent that supports MCP stdio can run the released package directly from GitHub:

uvx --from "crispasr-agent-transcriber[mcp] @ git+https://github.com/EmiyaKatuz/crispasr-agent-transcriber.git@v0.3.5" crispasr-agent-mcp

Use the same command and arguments in Claude Desktop, Cursor, or another MCP client. See AI agent integrations for a generic MCP configuration and Codex CLI command.

Maintainer publishing

End users do not need the release steps. Maintainers should follow the publishing guide for Codex Marketplace, PyPI, MCP Registry, and cross-agent distribution.

Required models

This tool does not download models automatically. Download these three GGUF files and keep them in a local directory (the repo's models/ folder works well):

The English F16 file is the highest-quality Cohere option and preserves the existing default. The same repository also provides smaller quantized files, including cohere-transcribe-q4_k.gguf; pass its exact local path if you choose that variant. All three upstream model families are Apache 2.0 licensed.

For automatic English/Chinese routing, pass both ASR paths. The language probe runs first, and only the matching model is loaded:

--english-model models\cohere-transcribe.gguf
--chinese-model models\qwen3-asr-1.7b-q4_k.gguf
--lid-backend firered --lid-model models\firered-lid-q2_k.gguf

For an explicit english or chinese profile, --model remains available as a single-model override.

CrispASR binary management

The tool auto-detects, installs, and updates the CrispASR binary from GitHub releases.

When --manage-server is set and no binary is found, it auto-installs before starting the server.

GPU detection

On install and update, the tool checks your hardware:

1. CUDA -- nvidia-smi available, or CUDA_PATH / CUDA_HOME set, or CUDA in PATH -> downloads crispasr-*-cuda variant.
2. Vulkan -- vulkaninfo or VULKAN_SDK set (only when CUDA is absent) -> downloads crispasr-*-vulkan variant.
3. CPU -- fallback when no GPU toolkit is detected.

macOS always uses the universal binary.

Profiles

auto mode runs FireRed language detection on the media, then routes English to Cohere or Chinese to Qwen3-1.7B. Mixed or uncertain content stops with a clear error asking you to re-run with --profile english or --profile chinese.

Usage

Managed server (tool starts CrispASR for you)

uv run python scripts/transcribe.py sample.wav `
  --profile auto `
  --manage-server `
  --english-model models\cohere-transcribe.gguf `
  --chinese-model models\qwen3-asr-1.7b-q4_k.gguf `
  --lid-backend firered --lid-model models\firered-lid-q2_k.gguf `
  --format srt `
  --out-dir outputs

Add --keep-server to leave the server running after transcription.

Manual server (you start CrispASR)

# Terminal 1 -- start the server
crispasr --server --backend cohere `
  -m models\cohere-transcribe.gguf `
  --port 8080

# Terminal 2 -- transcribe
uv run python scripts/transcribe.py sample.mp4 `
  --profile english `
  --server-url http://127.0.0.1:8080 `
  --format verbose_json

If the running server's backend doesn't match the selected profile, the tool prints the exact command you need to start the correct server.

Output formats

A .metadata.json sidecar is always written alongside the transcript.

Video files

Video files are detected automatically. ffmpeg extracts the audio track to a temporary mono 16 kHz WAV before sending it to CrispASR. The temporary file is deleted when transcription finishes.

All CLI flags

--profile auto|english|chinese
--format text|verbose_json|srt|vtt
--out-dir PATH
--server-url URL
--allow-remote-server
--manage-server
--keep-server
--model PATH               Local GGUF override for an explicit profile
--english-model PATH       Cohere model selected after English detection
--chinese-model PATH       Qwen3-ASR model selected after Chinese detection
--allow-model-auto-download
--lid-model PATH           Local LID model path
--lid-backend firered|silero|ecapa|whisper
--host HOST                Managed server host (default 127.0.0.1)
--port PORT                Managed server port (default 8080)
--language CODE            Language hint for transcription
--prompt TEXT              Initial prompt/context
--vad                      Enable voice activity detection
--diarize                  Enable speaker diarization
--diarize-method METHOD
--hotwords WORD,WORD       Comma-separated hotwords
--no-timestamps
--preprocess auto|always|never
--api-key KEY              If CRISPASR_API_KEYS is enabled
--crispasr-bin-dir PATH
--crispasr-bin PATH
--install-crispasr
--update-crispasr
--crispasr-status

MCP server

uv sync --extra mcp
uv run --extra mcp crispasr-agent-mcp

Exposed tools:

Security model

• No cloud uploads. Media files stay on the local filesystem.
• No remote servers by default. --server-url only accepts localhost unless --allow-remote-server is explicitly passed.
• No URL inputs. Only local file paths are accepted. URLs, S3, and other remote schemes are rejected.
• No shell injection. ffmpeg is called with argument lists and shell=False. No user-controlled strings are interpolated into shell commands.
• No model downloads by default. CrispASR model auto-download (-m auto) requires --allow-model-auto-download. The same guard applies to language detection models.
• Temporary files are cleaned up. Converted WAV files and LID probe windows are deleted when transcription finishes.
• Binary downloads are explicit. CrispASR binary installs only from the official CrispStrobe/CrispASR GitHub releases.
• Verified plugin releases. The npm installer requires the plugin ZIP to match the SHA-256 value published in the same GitHub Release.
• Narrow installer writes. The installer manages only its plugin directory and the named Personal marketplace entry. Updates preserve local models, binaries, and outputs.

Verify

uv run pytest
uv run ruff check .  # zero lint warnings

License

This project is licensed under the MIT License.

Third-party components and attribution

This tool orchestrates several independently-licensed projects. It does not bundle, fork, or redistribute their code -- it downloads pre-built binaries and calls them as subprocesses or HTTP services at runtime.

Model files must be downloaded separately by the user from their respective HuggingFace repositories. See Required models above.

Related projects

• CrispASR -- the ASR engine this tool wraps
• CrisperWeaver -- CrispASR's desktop GUI (not used by this tool)