kyutai-mcp

📖 Installation Guide — quick start, manual setup, and troubleshooting

Run Kyutai Moshi locally. Drive it from a glass dashboard, and plug the same surface into your agent stack via FastMCP.

Quick start Web UI Technical docs

FastMCP 3.1+ stdio + HTTP /mcp web 10924 / 10925 MCP HTTP 10926

---

Quick Start

git clone https://github.com/sandraschi/kyutai-mcp
cd kyutai-mcp
just

This opens an interactive dashboard showing all available commands. Run just bootstrap to install dependencies, then just serve or just dev to start.

Manual Setup

If you don't have just installed:

Web dashboard

Standard pages: Home, Actions, Tools (ops + MCP catalog), Apps (Glama / manifest), Moshi (and /talk), Status, Chat, Logger, Settings, Help.

The in-app Help page is the full operator manual (routes, Glom-On, APIs, troubleshooting).

---

MCP server (stdio)

uv run python -m kyutai_mcp

Wire this in Cursor/your client as a stdio MCP server. See docs/MCP.md.

---

Technical documentation

Doc	Contents
docs/WEBAPP.md	Ports, routes, stack, build
docs/MCP.md	Tools, transports, discovery
docs/MOSHI_SERVICE.md	Upstream Moshi process, HTTP probe
docs/GLOM.md	Local LLM attach (Ollama / LM Studio)
docs/VOICE_WORKFLOWS.md	Staged voice orchestration, prompts, examples, agentic boilerplates

---

Related: speech-mcp (cloud voice)

speech-mcp is the cloud counterpart to this project. Where kyutai-mcp runs Moshi locally on your GPU, speech-mcp connects to cloud speech APIs.

	kyutai-mcp (this repo)	speech-mcp
Engine	Moshi (Kyutai, open-source)	Gemini Live, Gemini TTS, Hume, ElevenLabs
Runs on	Local GPU (RTX 4090, CUDA)	Cloud APIs
Privacy	Fully offline	Cloud
Latency	Low (local)	Sub-second (Gemini Live)
Voice quality	Good	Very good to highest
Voice cloning	No	Yes (ElevenLabs IVC)
Multilingual	Limited	100+ languages (Gemini TTS)
Cost	Free after hardware	API usage costs

Use kyutai-mcp when privacy or offline operation matters. Use speech-mcp when voice quality, multilingual coverage, or voice cloning is the priority. Both expose the same portmanteau MCP tool pattern and can be run simultaneously on different ports.

---

This project adheres to SOTA 14.1 industrial standards for high-fidelity agentic orchestration:

• Python (Core): Ruff for linting and formatting. Zero-tolerance for print statements in core handlers (T201).
• Webapp (UI): Biome for sub-millisecond linting. Strict noConsoleLog enforcement.
• Protocol Compliance: Hardened stdout/stderr isolation to ensure crash-resistant JSON-RPC communication.
• Automation: Justfile recipes for all fleet operations (just lint, just fix, just dev).
• Security: Automated audits via bandit and safety.

License

See repository LICENSE if present; Kyutai/Moshi upstream has its own licenseconsult Kyutai Moshi for model and server terms.