MCP and agents

Let your agent understand recordings and videos

Register EdgeSpeak once over MCP and agents like Claude Code can transcribe, align and generate speech on your computer.

Direct answer

EdgeSpeak provides 13 MCP tools for transcription, speaker diarization, forced alignment, sentence splitting, text to speech, voice management and local language models. After `claude mcp add edgespeak -- edgespeak-cli mcp`, your agent can call them in plain language, and the speech processing runs on your computer.

Connect with one command

Install and activate edgespeak-cli, then register EdgeSpeak in Claude Code. Other MCP clients that support stdio use the same command. Over stdio, MCP reuses the models loaded by the running desktop app. Without the app, run edgespeak-cli serve and connect to its HTTP /mcp endpoint, which runs the bundled local engine itself.

claude mcp add edgespeak -- edgespeak-cli mcp

What the 13 tools cover

Grouped by job:

  • Transcription and speakers: edgespeak_transcribe_file, edgespeak_transcribe, edgespeak_diarize_file
  • Alignment and sentences: edgespeak_align, edgespeak_segment_sentences
  • Speech and voices: edgespeak_create_speech, edgespeak_list_voices, edgespeak_add_voice, edgespeak_delete_voice, edgespeak_speech_examples
  • Models and usage: edgespeak_list_models, edgespeak_create_response, edgespeak_usage_summary

A full example: editing video with Claude Code

Claude Code already drives FFmpeg; EdgeSpeak tells it when every word starts and ends. Ask it to cut pauses longer than two seconds from an interview and add word-by-word subtitles: the agent transcribes the file with word timestamps, builds the FFmpeg cuts from those times, then exports subtitles and burns them in with the EdgeSpeak Agent Skill.

Where your data goes

Transcription, alignment and speech generation run on your computer with local models. The text a tool returns goes to your agent and is handled by the model the agent uses; if that is a cloud model, that text is sent to its service.

MCP, CLI or Agent Skill

MCP lets an agent call individual tools. The Agent Skill teaches an agent the full workflow with the CLI: transcribe, split sentences, align and export subtitles. Scripts and automation can call the CLI or the local API directly. You can use all three together.

FAQ

Which MCP clients work?

Any MCP client that supports stdio, such as Claude Code. You can also use the HTTP /mcp endpoint of the desktop app or the headless service.

Does the desktop app need to be open?

For stdio, yes: the stdio server connects to the running app and returns could not reach gateway while it is closed. Without the app, run edgespeak-cli serve and use its HTTP /mcp endpoint, or call one-shot edgespeak-cli commands such as transcribe or align.

Which tools change files?

Only adding a voice, deleting a voice and generating speech write managed local files. The other 10 tools are read-only.

Evidence links