Earshot MCP Server: Setup for Claude, Cursor and Other Apps

• Measures loudness, peaks and timing of a file

  Gives exact meter readings such as duration, loudness in LUFS and true peak, without AI models. The README says to trust these numbers.

  Try: “How loud is this audio file?”

  Read · measure

• Says what a sound is

  Gives tags for the source, material, action and character of a sound using two local AI models. The first call takes 20 to 60 seconds while they load.

  Try: “What does this sound clip contain?”

  Read · describe

• Checks if a sound fits a description

  Scores how well one sound matches a text description and gives a verdict. Give alternative wordings so it can say which fits best.

  Try: “Does this clip sound like a door slam?”

  Read · match

• Ranks many sounds against a description

  Sorts candidate files by how well they match a description, to pick the best one.

  Try: “Which of these files is the best whoosh?”

  Read · rank

• Finds moments inside a long recording

  Searches one long recording for segments that match a description and returns start and end times.

  Try: “Find where the applause starts in this recording.”

  Read · find

• Compares a sound to reference recordings

  Finds the nearest references and says how typical the sound is among them.

  Try: “Is this kick like my reference kicks?”

  Read · compare

• Compares two sounds side by side

  Says which of two sounds is closer to a description or a reference set, and how they differ in measured values.

  Try: “Is version B better than version A?”

  Read · ab

• Judges real recording versus synthesized

  Estimates whether audio sounds like a real recording or like simple code-made sound.

  Try: “Does this sound like a real recording?”

  Read · realness

• Listens through a whole soundtrack

  Produces a transcript of events with time, level and tags for a soundtrack or film.

  Try: “Give me a timeline of the sounds in this film.”

  Read · timeline

• Draws a picture of a soundtrack

  Makes an annotated image of a soundtrack or film on one time axis, with frames, sections and labelled events, and saves it as a file.

  Try: “Draw a sheet of this film's sound.”

  Write · sheet

• Scores a launch-film soundtrack

  Gives a scorecard of loudness, peak, dynamics and spectral balance for a launch film's sound.

  Try: “Critique the sound of this launch film.”

  Read · critique

• Checks the earshot install

  Checks packages, ffmpeg and the models, and optionally loads them to time one run.

  Try: “Is earshot set up correctly?”

  Read · doctor

• Sends audio to Gemini for an opinion

  Uploads audio to Google Gemini for a written critique. It needs a Gemini key in the server environment and works only after you agree.

  Try: “Ask Gemini for a critique of this track.”

  Network · llm_critique

• Lists the shipped sound bank

  Lists the sound effects that ship with earshot, with length, peak time and the cue kinds they suit. It uses no AI models.

  Try: “What whoosh sounds are in the bank?”

  Read · sfx_list

• Renders one bank sound to a WAV file

  Renders a sound from the bank to a WAV file, adjusted to the picture, in about a second.

  Try: “Render a crisp whoosh of 0.4 seconds to a WAV.”

  Write · sfx_render

• Renders a film's sound kit

  Renders the takes for every role in a film's kit.json into WAV files, plus a lock file.

  Try: “Render the sound kit for this film.”

  Write · kit_render

Watch out for

• The llm_critique tool uploads audio to Google Gemini

  The llm_critique tool sends audio to Gemini, needs your GEMINI_API_KEY, and asks for your explicit agreement first (a consent flag). Leave the key unset if you never want that. github.com

• It writes files to paths you give it

  Three tools write WAV or image files to an absolute path. Check where the assistant saves them. github.com

• Large first download

  The first use downloads about 1.1 GB of models into the repo folder. The README says it never contacts the Hub afterwards. github.com

Security advisories