🪨 Caveman MCP — Me Smash Text, Make Small

TL;DR: Feed URL, file, image, diff, log. Get Wenyan. Big text become tiny text. Code block survive. Number survive. Everything else get crushed by rock.

Two-pass compression: mechanical filler-word drop → LLM Wenyan classical Chinese conversion. Any OpenAI-compatible endpoint (OpenRouter, MiniMax, etc).

---

Tools

condense_url

Fetch any URL → extract main content → compress.

Smart routing by URL pattern:

URL	Extractor
youtube.com/watch, youtu.be/	Transcript (VTT captions)
github.com/{owner}/{repo}	README + file tree + activity
github.com/{owner}/{repo}/pull/{n}	PR diff + description
arxiv.org/abs/{id}	Abstract + section headings
news.ycombinator.com/item?id=	Top 20 comments (HN-ranked)
reddit.com/r/*/comments/*	Top 20 comments by score
RSS/Atom feed	Batch item summaries
Anything else	Enhanced HTML → markdown pipeline

condense_file

Read any local file → extract text → compress.

Extension	Extractor
.png .jpg .jpeg .webp .gif	Vision LLM → wenyan description
.mp3 .wav .m4a .ogg .flac	Whisper API → transcript → compress
.pptx	Slide text + speaker notes
.xlsx .csv	Column stats + top values summary
.pdf	Text extraction
.docx	XML text extraction
.md .txt .html .htm .rst	Raw text / readability

condense_git

Compress git artifacts. One of text, path, or pr_url required.

Auto-detects format: unified diff → per-file summaries, git log → topic clusters, git blame → authorship summary, GitHub PR URL → fetch + compress.

condense_log

Compress error logs and stack traces.

Detects language (Go / Python / JS / Java / Rust), strips stdlib frames, deduplicates repeated errors with ×N count, preserves exception type + message + file:line exact.

condense_text

Raw text in. Wenyan out. No fetching.

---

Compression Pipeline

input text
  → mechanical pass: drop articles (a/an/the), fillers (just/basically/actually/...),
                     preserve code blocks, URLs, identifiers, numbers
  → LLM Wenyan pass: classical Chinese ultra-compression
      HTML/structured content uses differential pressure:
        h1/h2  → light compression (landmark)
        body   → full wenyan prose
        aside  → ultra-compress or omit
        nav/ad → drop
  → result: {compressed, original_chars, compressed_chars, ratio, method}

Typical ratios: 65–80% of original. Code blocks survive intact.

---

Setup

Claude Code (recommended)

{
  "mcpServers": {
    "caveman": {
      "command": "npx",
      "args": ["-y", "@standardbeagle/caveman-mcp"],
      "env": {
        "LLM_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

No install step — npx downloads and runs on first use.

Go install

go install github.com/standardbeagle/caveman-mcp@latest

Build from source

git clone https://github.com/standardbeagle/caveman-mcp
cd caveman-mcp && go build -o caveman-mcp .

---

Examples

Condense a webpage:

{"name": "condense_url", "arguments": {"url": "https://modelcontextprotocol.io/introduction"}}

MCP（模型語境協議）者，開源標準也，用以聯接AI應用與外部系統...
ratio: 77.1% | 1794 → 1384 chars | mechanical+llm

Condense a git diff:

{"name": "condense_git", "arguments": {"path": "/tmp/changes.diff"}}

Compress logs:

{"name": "condense_log", "arguments": {"text": "panic: runtime error: index out of range...\ngoroutine 1 [running]:\nmain.process(...)..."}}

Image description:

{"name": "condense_file", "arguments": {"path": "/tmp/architecture-diagram.png"}}

---

Optional Dependencies

• yt-dlp — fallback for YouTube videos without auto-captions (pip install yt-dlp)

---

MCP SDK

Built on the official MCP Go SDK v1.5.0. Stdio transport.