cliproxy-media

Source: https://github.com/bencoremans/site/tree/main/skills/cliproxy-media

Analyze images and PDFs via CLIProxyAPI (Claude Max subscription, zero extra cost).

Setup

Set the endpoint to your CLIProxy instance:

CODEBLOCK0

For Docker setups, replace your-host with your container hostname (e.g. cliproxyapi, localhost, or the container IP).

Quick start

CODEBLOCK1

What works ✅ / What doesn't ❌

✅ Supported file types

Multiple files at once: Provide multiple paths before the question. Max ~100 per request (Anthropic limit).

❌ Not supported

• - Office files (.docx, .xlsx, .pptx) — Workaround: convert to PDF
• Audio (.mp3, .wav, .ogg) — Use Whisper for transcription
• Video (.mp4, .mov, .avi) — Not supported by the model
• Other document types (.txt, .html, .md as document) — Send text directly as a string

⚠️ System prompt warning

CLIProxyAPI accepts only the array notation for system prompts. The string notation is silently ignored — the model does not see it, but you also won't get an error message!

CODEBLOCK2

The --system argument in analyze.py automatically uses the correct array notation.

Configuration (env vars)

Example:
CODEBLOCK3

Additional options

CODEBLOCK4

Compatibility

This script works with any API that supports the Anthropic Messages format:

Note for non-CLIProxy endpoints: Some proxies do accept string notation for system prompts. Always use array notation for maximum compatibility.

Known limitations of CLIProxyAPI

• - temperature and top_p may not be used at the same time (HTTP 400)
• PDF as document with URL source does not work (Unable to download the file)
• Only claude-sonnet-4-6 and claude-opus-4-6 available (haiku is deprecated)
• INLINECODE37 is always not_available in the response

Direct Python API

If you want to call the script from your own Python code:

CODEBLOCK5

Or use the built-in exec tool:
CODEBLOCK6