Perception MCP server

Read from the package source without running it. The installed server may list more.

• analyze_audio

  Full one-call audio report: loudness/quality, tempo, musical key and pitch. The audio is loaded once and every analysis runs on it.

• analyze_loudness

  Measure loudness and technical quality of a recording — the "will platforms accept this?" check for podcasts, voice-overs and mixes.

• check_contrast

  Audit text contrast in a screenshot against WCAG 2.x: OCR finds the text, the exact W3C relative-luminance formulas score each line, and every line gets a pass/fail for AA and AAA.

• check_mesh

  QC a 3D mesh for printability and geometric hygiene: watertightness, winding consistency, degenerate/duplicate faces, volume, extents — with a printable yes/no and the reasons.

• check_video

  QC a video for the defects that make footage unusable: black stretches, frozen picture, dead audio, soft focus — with exact timestamps for each finding.

• compare_audio

  Compare two recordings: tuning difference in cents, tempo difference, harmonic (key) compatibility for DJ mixing, and loudness difference.

• compare_images

  Compare two images pixel-exactly and structurally — the "did my change do what I intended and nothing else?" tool for screenshots, renders, and any before/after check.

• compare_mesh

  Measure how far two 3D models diverge geometrically: ICP alignment, then two-sided sampled surface distances (chamfer mean + Hausdorff max), plus an overlay render (A grey, B translucent red).

• compare_videos

  Diff two cuts of a video at the SCENE level: which shots are identical, changed, added or removed — robust to timeline shifts that break naive frame-by-frame comparison.

• describe_audio

  Describe the structure of a recording in words: the timeline is split into sections by sound character, each labeled with loudness, brightness, rhythm density and texture, plus an overall guess whether it's music, speech or something else.

• describe_video

  Understand a whole video in ONE call: stream facts, shot structure, audio profile, spoken text, and a contact sheet with a timestamped frame from every scene.

• detect_key

  Detect the musical key (tonality) of a recording, e.g. "A minor".

• detect_pitch

  Detect the fundamental pitch of a monophonic audio recording (single voice, string, whistle, tuning fork, etc.).

• detect_scenes

  Detect scene cuts in a video and return their timestamps — a shot-by-shot table of contents you can then sample with get_frames.

• detect_silence

  Map silent gaps in a recording with timecodes — useful for finding pauses to trim in podcasts, voice-overs or interviews, and for checking leading/trailing silence.

• estimate_tempo

  Estimate the tempo (BPM) of a piece of music.

• extract_text_boxes

  Read printed text from an image WITH exact pixel coordinates for every line — what plain vision can't give you: where the text is, not just what it says.

• find_media

  Find recent media files on the USER'S machine when you have no path — the user attached a file in chat (attachments give you rendered pixels, not bytes; images carry no filename at all) or says "the video I just downloaded". Do not look for the user's files with your own file/code tools: those see only your cloud sandbox, never the user's disk — this tool is the one that runs on their machine. The

• get_frames

  SEE a video: pull the exact frames you need as images, each stamped with its timecode.

• get_metadata

  Read file metadata without decoding the audio: duration, format, sample rate, channels, bitrate, file size, and embedded tags (title/artist/album/date/genre).

• get_youtube_transcript

  Fetch the transcript (captions/subtitles) of a YouTube or other yt-dlp-supported video as text with timestamps — WITHOUT downloading the audio. This is the fastest way to know what is said in a video.

• measure_image

  Measure an image with numbers a vision model cannot extract by looking: exact pixel colors at given points, pixel distances between points, image dimensions, and the dominant color palette.

• media_assert

  Verify a media file against declarative expectations and get a pass/fail report per check — a test harness for generated or processed media. Use this after producing/editing a file ("did the export come out right?") instead of reading raw analysis numbers, or before publishing ("will platforms accept this?").

• prepare_text_tiles

  Make hard-to-read text READABLE: small print, dense screenshots, photos of documents, handwriting. The page is cleaned gently (illumination flattening, deskew) and returned as a few big overlapping blocks, cut only at natural gaps between text lines — then read the blocks yourself.

• render_font_specimen

  SEE a font and verify what it supports: a rendered specimen sheet plus hard facts from the font tables — the only way to check a typeface before shipping it.

• render_model

  SEE a 3D model: seven orthographic views (front / right / back / left / top / bottom / isometric) on one sheet, with the bounding-box extents in the caption.

• render_spectrogram

  Render a mel spectrogram of the audio as a PNG image, so the frequency content over time can be SEEN: melodies, drops, vocals, noise bursts, silence, filter sweeps.

• render_waveform

  Render the waveform plus an RMS loudness curve as a PNG image — a visual overview of dynamics: where it's loud, quiet, silent, or clipped.

• transcribe_audio

  Transcribe speech in an audio recording to text with timestamps, locally via Whisper (faster-whisper, CPU).

• video_dynamics

  Measure HOW a video moves: cut rhythm, motion energy and loudness on one timeline, plus hook metrics for the opening seconds — the raw ingredients of "does it hold attention".

• video_heatmap

  See WHERE things happen in the frame: activity maps drawn over a real frame, with a 3x3 grid breakdown.

• view_media

  SEE a file — one call renders the whole image (or one PDF page), automatically downscaled and compressed to fit any client's tool-result size limit. Use this FIRST whenever the user refers to a picture or document you have not seen: view_media for the full view, then zoom_grid only for small details. Do NOT try to view a whole image through zoom_grid — it magnifies and will overflow the result lim

• zoom_grid

  Magnify a region of an image and overlay a coordinate grid — a microscope for details too small to see at full size: tiny text, 1-px borders, icon pixels, exact element positions. (To simply LOOK at a whole image, use view_media instead — magnifying a full image overflows the response budget.)

• Code scan50 source files scanned25/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitylocal package, no credentials required12/15
• Maintenancerepository not readable: repo not found3/15
• Maintainer identityno repository or website to verify2/10

Perception: common questions

Is Perception MCP server safe?

With care: it is graded C, so read the findings first (65/100). Read the Perception safety report

How do I install Perception?

It runs on your machine. Copy the Claude Code, Claude Desktop or Cursor config from the install section.

Does Perception need an API key?

Not as far as the registry entry and our scan can tell: no credentials are declared or required.

Is Perception maintained?

The latest release is v0.1.7.

What can I use instead of Perception?

Servers from other publishers that do the same job: anymd MCP server, SuperCMO MCP server and Rendobar MCP server. Compare all Perception alternatives.