Best for
- Users who need local or private audio transcription
- Users already working with Docker, Claude, or Cursor
- Workflows that require FunASR or SenseVoice models
F
Funasr MCP
Official
Private local audio transcription with FunASR.
- Category
- Other #101 of 230
- Stars
- ★ 20.7k Very popular
- Transport
- stdio (local process)
- Runtime
- Docker
- Credentials
- No credential needed
- License
- MIT
- Last commit
56FMRS · C
This is an upstream FunASR MCP server manifest, version 0.1.2, distributed as a Docker package over stdio for private local audio transcription. The manifest declares no tool names, so tool-level capabilities and invocation parameters cannot be verified further.
Strongest · Maintenance 14/20 Weakest · Reliability 9/20
Why each score
- Reliability 9/20
- Static evidence shows a versioned OCI Docker package, stdio transport, and an explicit audio-directory mount, making the normal startup path plausible. However, the MCP server source, tool list, initialization/tool-list handshake evidence, tests, and CI are not provided; error handling, dependency failures, and actual tool behavior cannot be verified, so points are deducted.
- Security and permissions 11/20
- The audio directory is mounted read-only, model caching is isolated in a persistent volume, and the server is described as using local inference; no credentials, deletion, payment, deployment, or remote-command defaults or real secrets are shown. However, models are downloaded externally, while image provenance, network data flows, tool permission boundaries, sensitive-audio handling, and confirmation mechanisms are not sufficiently documented, so points are deducted.
- Maintenance 14/20
- The repository is not archived, the MIT license is explicitly stated, and the maintainer/repository identity is clear. The README lists continuous PyPI, runtime, and feature releases through 2026-07-24, indicating recent maintenance. There is no verifiable commit history, CI evidence, vulnerability-response channel, dependency-update policy, or explicit support commitment, so full marks are not justified.
- Documentation 12/20
- The README covers installation, Docker mounts, model selection, platform limitations, deployment methods, model-license distinctions, and several usage examples. MCP-specific evidence for tool parameters, return structures, error handling, client configuration, resource limits, and privacy/network behavior is missing, and the supplied material cannot verify that implementation matches the documentation, so points are deducted.
- Setup experience 10/20
- The manifest clearly specifies an OCI image, stdio transport, required audio-directory mount, and model-cache volume, which should be connectable from a stdio-capable client. It lacks concrete Claude/Cursor client configuration, startup logs, or connection verification; Docker, host-path configuration, and model downloads are required, while platform compatibility and first-run failure handling remain unclear, so points are deducted.
Static review · not runListed 2026-07-29
Read the FMRS scoring method →Fit and risk
What it can accessReads local files
Not for
- Integrations requiring a declared tool list or precise MCP tool contract
- Users seeking a cloud-hosted transcription service
- Users assuming the MIT code license applies to all pretrained model weights
Required permissions
- Read access to audio files in the mounted host directory
- Write access to the ModelScope model-cache volume for downloaded files
- Access to audio paths under /audio inside the container
Risks and side effects
- The mounted host audio directory is readable by the container, although the manifest specifies a read-only mount
- Model files are downloaded and retained in the ModelScope cache volume
- The FunASR toolkit is MIT-licensed, but pretrained model weights may have separate licenses shown on their model cards
- Because no tool names are declared, the available operations cannot be confirmed from the manifest
Setup
Before you start
Runtime：Docker
Run ghcr.io/modelscope/funasr-mcp:0.1.2 with Docker. Mount the host directory containing audio files read-only at /audio and use /audio/<file> when invoking the server. Mount a Docker volume named funasr-mcp-cache at /root/.cache/modelscope to persist downloaded ModelScope model files.
Check that it works
After connecting the client to the server, ask it to transcribe an audio file under the mounted /audio directory; a returned transcript confirms the MCP server is working.
Troubleshooting
- Verify that the host audio directory is mounted at /audio
- Use /audio/<file> rather than the host filesystem path when calling the server
- Confirm that Docker is using image version 0.1.2
- Check that the funasr-mcp-cache volume is mounted correctly so downloaded model files can be reused
- Confirm that the audio file exists inside the mounted directory
Things to try
Once connected, you can ask your AI assistant things like:
- Transcribe /audio/sample.wav to text
- Run speech recognition on the recording in my mounted audio folder
- Transcribe /audio/meeting.wav and include timestamps
Use cases
Transcribing local audio files
Connecting FunASR speech recognition to Claude or Cursor
Running on-device inference without sending audio to a cloud service
Supported clients
ClaudePartial support
CursorPartial support
Listed from the project's documentation, not tested by this site.
Overview
FunASR MCP Server is provided by the FunASR project for local audio transcription with FunASR and SenseVoice. The broader FunASR toolkit also supports streaming ASR, VAD, punctuation, speaker diarization, emotion recognition, and audio-event models, but this server manifest does not declare specific tool names. The server uses a Docker OCI image with stdio transport.
Similar servers
Hevy MCP
77 · B
Manage your Hevy workout data from AI assistants via MCP.
Firecrawl MCP Server
75 · B
Firecrawl's official MCP server for web search, scraping, and structured extraction for AI agents
BioMCP
75 · B
One binary. One grammar. Evidence from the biomedical sources you already trust.
Compartment
75 · B
Encrypted, fully offline long-term memory for AI agents
Source revision 879a219c7171 Data synced 2026-10-19 Read the FMRS scoring method