A local FastMCP stdio server for general-purpose FunASR speech processing.

• Stdio-native for local MCP clients ([mcp_servers.*]).
• Audio + video support (ffmpeg extraction for video input).
• ASR + VAD + async task tools in one server.
• Automatic backend selection (env override > cuda > mps > vulkan(if torch-supported) > cpu).
• Runtime model management and cache control.

• asr_detect_backend()
• asr_list_models(kind='all')
• asr_load_model(model='', backend='auto', force_reload=False, model_load_kwargs=None)
• asr_load_vad_model(model='', backend='auto', force_reload=False, model_load_kwargs=None)
• asr_clear_model_cache(kind='all')
• asr_validate_media(path)
• asr_transcribe_audio(path, language='auto', timestamps=True, model='', backend='auto', generate_kwargs=None, model_load_kwargs=None, include_raw=False)
• asr_transcribe_video(path, language='auto', timestamps=True, model='', backend='auto', generate_kwargs=None, model_load_kwargs=None, include_raw=False)
• asr_start_transcription_task(path, media_type='auto', language='auto', timestamps=True, model='', backend='auto', generate_kwargs=None, model_load_kwargs=None)
• asr_get_task_status(task_id)
• asr_get_task_result(task_id)
• asr_list_tasks(status='all', limit=20)
• asr_cancel_task(task_id)
• asr_delete_task(task_id)
• asr_detect_speech_segments(path, media_type='auto', model='', backend='auto', generate_kwargs=None, model_load_kwargs=None, include_raw=False)

• Python 3.10+
• uv
• ffmpeg in PATH

uv sync

• FUNASR_BACKEND: force backend (cuda|mps|vulkan|cpu)
• FUNASR_MODEL: default ASR model alias/name
• FUNASR_VAD_MODEL: default VAD model alias/name
• FUNASR_TASK_WORKERS: async task worker count (default 2)

uv run funasr-mcp

[mcp_servers.funasr]
command = "uv"
args = ["run", "--project", "D:/path/to/mcp-funasr-stdio", "funasr-mcp"]
enabled = true
startup_timeout_sec = 30
tool_timeout_sec = 120

1. asr_start_transcription_task(...)
2. Poll asr_get_task_status(task_id)
3. Fetch asr_get_task_result(task_id) when status=completed
4. Optionally asr_delete_task(task_id) to release memory

uv run python -m py_compile src/funasr_mcp/*.py