FunASR MCP Server
Model Context Protocol server that gives AI assistants local audio transcription with SenseVoiceSmall by default.
Setup
1. Install dependencies
2. Optional: run with Docker
The Dockerfile starts the MCP server over stdio and is suitable for MCP directory
checks that initialize the server and call tools/list.
docker build -t funasr-mcp examples/mcp_server
docker run --rm -i \
-e FUNASR_DEVICE=cpu \
--mount type=bind,src=/path/to/audio,dst=/audio,readonly \
--mount type=volume,src=funasr-mcp-cache,dst=/root/.cache/modelscope \
funasr-mcp
Verify the image entrypoint and MCP handshake without downloading a model:
python examples/mcp_server/smoke_test.py funasr-mcp
When submitting this server to MCP directories such as Glama, use this folder as
the Docker build context so the container entrypoint runs funasr_mcp.py.
The repository root glama.json declares GitHub maintainer ownership for Glama,
while the glama.json file in this directory declares the container command and
metadata for directory scanners.
Official MCP Registry
The versioned Registry metadata is in server.json. It points to
the public GHCR image and asks clients to mount one host audio directory at
/audio read-only. The Dockerfile carries the matching OCI ownership label:
Current published release:
LABEL io.modelcontextprotocol.server.name="io.github.modelscope/funasr-mcp"
The release workflow validates the metadata against the official schema, builds
the image, performs an MCP initialize and tools/list handshake, pushes the
versioned image to GHCR, and publishes server.json through the official
mcp-publisher CLI.
To release a new MCP server version:
- Update
version and the OCI image tag in server.json together.
- Merge the change after the MCP validation workflow passes.
- Have a
modelscope organization Owner push the matching mcp-v<version>
tag, for example mcp-v0.1.2.
- Approve the protected
mcp-registry-publish environment deployment.
The official Registry only grants the io.github.modelscope/* namespace to a
GitHub organization Owner. Keep the publish environment restricted to release
tags and require a maintainer approval because its OIDC token can publish the
organization namespace.
After publication, clients can run the pinned image directly:
docker run --rm -i \
--mount type=bind,src=/path/to/audio,dst=/audio,readonly \
--mount type=volume,src=funasr-mcp-cache,dst=/root/.cache/modelscope \
ghcr.io/modelscope/funasr-mcp:0.1.2
When using the container, pass tool paths under /audio, such as
/audio/meeting.wav.
Glama submission checklist
Use these values when adding the server at https://glama.ai/mcp/servers:
| Field | Value |
|---|
| Repository URL | https://github.com/modelscope/FunASR |
| Docker build context | examples/mcp_server |
| Dockerfile path | examples/mcp_server/Dockerfile |
| Server command | python /app/funasr_mcp.py |
| Expected MCP tool | transcribe_audio |
After Glama finishes evaluation, verify that the listing or API shows an
evaluated quality score before adding it to directory PRs. An HTTP 200 from the
listing or badge endpoint alone does not prove that a score has been assigned:
[](https://glama.ai/mcp/servers/modelscope/FunASR)
If the badge endpoint still returns 404, keep the badge out of external
directory submissions until the Glama listing is live.
Directory listings
The FunASR MCP server is published in the official Registry and listed on mcp.so:
Claude Code (~/.claude.json):
{
"mcpServers": {
"funasr": {
"command": "python",
"args": ["/path/to/examples/mcp_server/funasr_mcp.py"],
"env": {"FUNASR_DEVICE": "cuda"}
}
}
}
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"funasr": {
"command": "python",
"args": ["/path/to/funasr_mcp.py"],
"env": {"FUNASR_DEVICE": "cpu"}
}
}
}
Cursor (Settings → MCP Servers → Add):
- Command:
python /path/to/funasr_mcp.py
- Environment:
FUNASR_DEVICE=cuda
transcribe_audio
Transcribe a speech audio file to text.
Parameters:
| Name | Type | Required | Description |
|---|
audio_path | string | Yes | Path to audio file (wav, mp3, flac, m4a, ogg) |
language | string | No | auto, zh, yue, en, ja, or ko (default: auto) |
Returns: Transcribed text with per-segment timestamps when the model returns them.
Example Usage
Once configured, ask your AI assistant:
- "Transcribe the meeting recording at ~/Downloads/meeting.wav"
- "What was said in this audio file? /path/to/interview.mp3"
- "Convert this voice memo to text: ~/voice_note.m4a"
Environment Variables
| Variable | Default | Description |
|---|
FUNASR_DEVICE | cpu | Device: cuda, cpu, or mps |
FUNASR_MODEL | iic/SenseVoiceSmall | Model name or local model path passed to AutoModel |
Features
- Five-language transcription — Mandarin, Cantonese, English, Japanese, and Korean
- Automatic detection or explicit hints —
auto, zh, yue, en, ja, and ko
- VAD segmentation — splits longer audio before recognition
- Optional segment timestamps — included only when the configured model returns them
- Configurable local inference — choose the model and CPU, CUDA, or MPS with environment variables
- No API key needed — fully local inference
- MIT licensed, privacy-friendly (audio never leaves your machine)
Verified Compatibility
| Tool | Status |
|---|
| Claude Code | ✅ Tested |
| Claude Desktop | ✅ Compatible |
| Cursor | ✅ Compatible |
| Windsurf | ✅ Compatible |
| Any MCP client | ✅ Standard protocol |