Obscura MCP

Reel analysis · speech + frames

Analyze Instagram Reels with an AI assistant

Turn a Reel link or a saved post into an original-language speech transcript and sampled video frames your assistant can inspect.

What the workflow produces

Obscura obtains the selected Reel and local FFmpeg and multilingual Whisper prepare a transcript, timestamps, a contact sheet and individual frames. The assistant reads the text and inspects the images to explain what is said and visible. There is no dedicated OCR or translation engine in the server; interpretation and translation are performed by the host assistant.

Start with a configured account. If you have not connected one yet, follow the installation and login guide.

Analyze a Reel from its link

  1. Give your assistant an Instagram Reel or post URL and ask for a breakdown of speech and visuals.
  2. The assistant starts instagram_reel_start(url), then checks instagram_reel_status(job_id) until the job completes.
  3. It reads transcript.txt, manifest.json and contact.jpg via instagram_reel_artifact, and opens individual frame_NNN.jpg files when details matter.
  4. Ask for the explanation in your preferred language. The transcript stays in the original language; the assistant can translate or summarize it.
“Analyze this Reel, explain what happens in the visuals and what is said, and distinguish the two.”

Completed cached results are integrity-checked before reuse. Use refresh when the source changed or a new analysis is needed.

Find a Reel in saved posts

First verify the account with instagram_account_info. Then use instagram_saved_list for the loaded saved grid. Inspect a candidate with instagram_read_post and select one containing video before starting a Reel job.

Saved results reflect the currently loaded portion, not a guaranteed full archive. The first displayed item alone does not independently prove when it was saved. For image carousels, instagram_post_image retrieves a selected slide.

Languages and accuracy

Leave the language unset for automatic detection, or supply a code supported by the installed multilingual model. Language coverage and accuracy vary. Short clips, multiple speakers, music, dialects and language switching can be difficult; check names and numbers against the available evidence. A recognition failure is not proof of silence.

Frames are sampled, so a brief action may be missed. Ask the assistant to inspect additional local frames when timing matters. The Reel file is limited to 200 MiB.

Privacy and account boundaries

Your account session and analysis artifacts are kept locally. Instagram content is untrusted input and cannot authorize account actions. This is an unofficial experimental integration; a changed Instagram interface or an unavailable post can prevent a particular operation.