Skip to main content
Glama

stts

stts is a voice window for coding agents: an MCP server (stts-mcp, tools stt and tts) talks to a small local daemon on 127.0.0.1:15986, which drives a Chrome app window that listens with on-device speech recognition and speaks with Piper. It ships as the Claude Code plugin stts from the stts-marketplace and also runs over stdio behind an MCP gateway. Data lives in %LOCALAPPDATA%\stts\. MIT licence, see LICENSE.

Install

  1. /plugin marketplace add markkennethbadilla/stts

  2. /plugin install stts@stts-marketplace

  3. Run /stts to start a voice conversation.

Needs Node 20+, Chrome and Python 3.10+ on PATH. On the first start the daemon fetches the pinned voice and speech parts (Piper, the en_GB-jenny_dioco-medium voice, onnx-asr with NVIDIA Parakeet TDT 0.6B v3, sherpa-onnx's Silero voice detector; about 760 MB, SHA256-checked; about 20 minutes on a typical connection, measured at 0.5 MB/s from Hugging Face) into %LOCALAPPDATA%\stts\ (~/.local/share/stts/ on Linux) and logs each step to daemon.log there. Until it finishes, the window speaks with the browser's voice. To run it yourself: node scripts/install.mjs in the plugin folder (--check lists what is missing).

Models folder: STTS_MODELS_DIR, else modelsDir in <data>/config.json, else <data>/models. Point it at a folder you already keep models in to share one copy.

No npm install or build happens on install: the bundled dist/ (self-contained dist/mcp.js and dist/daemon.js, plus the page) is committed to the repository. After changing src/, web/ or public/, run npm run build and commit dist/ with the change; the push gate fails on a stale dist/.

Gateway (MCPJungle, every agent): register a stdio server with command node <repo>/dist/mcp.js (no build needed, dist/ is committed) and env STTS_WHO=agent, so a helper agent cannot close the session's window. Our own machines register it from mkb-agentops (spec 010).

For development: npm ci, then npm run check and npm run build. npm run e2e builds, starts the daemon on STTS_TEST_PORT (default 15990) with its own data dir under test-results/, and drives the page in headless Chromium with a fake speech recogniser, fake media, page.clock and a fake Piper server. Locally it uses the shared Playwright headless shell from mkb-agentops/versions.json.

Checks: a local Trunk push gate (npm run gate, spec 011), no GitHub Actions. Run trunk git-hooks sync once per clone.

Related MCP server: VoxMesh

Specs

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables bidirectional voice interaction for Claude Code using local speech-to-text and text-to-speech models optimized for Apple Silicon. It provides tools to listen to user speech via microphone and speak responses aloud through system speakers.
      16
      Apache 2.0
    • A
      license
      Not graded
      quality
      C
      maintenance
      Adds voice conversation capabilities to AI agents via MCP, enabling local speech recognition and synthesis with tools like speak, listen, and ask_by_voice for interactive voice interactions.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to access local voice synthesis, zero-shot cloning, and voice catalog tools through native MCP tool calls for applications like Claude and Cursor.
      37 npm
      AGPL 3.0