(V)(;,,;)(V)

Mr. Krab

The Token-Pinching Crustacean for OpenClaw

"I like tokens the way I like me money... SAVED."

---

Mr. Krab hooks into OpenClaw's prompt lifecycle and pinches every wasted token from your LLM calls. He's got two claws — one for conversation history, one for tool definitions — and he uses them on every single call.

He's query-aware: ask about GitHub issues and he snips away the Spotify docs, the deployment chatter, and that Slack conversation from 3 hours ago. Only what matters gets through. Everything else? snip snip.

Powered by ScaleDown's compression API. Falls back gracefully on failure — Mr. Krab is cheap, not reckless.

---

See Him in Action

SCALEDOWN_API_KEY="your-key" npx tsx demo.ts

      ___           ___
     /  /\         /  /\
    /  /::|       /  /::|
   /  /:|:|      /  /:|:|
  /  /:/|:|__   /  /:/|:|__
 /__/:/ |:| /\ /__/:/ |:| /\
 \__\/  |:|/:/ \__\/  |:|/:/
     |  |:/:/      |  |:/:/
     |__|::/       |__|::/
      \__\/         \__\/

  MR. KRAB x OpenClaw
  "I like tokens the way I like me money...
   SAVED."

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  (V)(;,,;)(V)  Mr. Krab's Token-Pinching Demo for OpenClaw
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  The Prompt (before Mr. Krab gets his claws on it):
    Conversation:  15 messages  (~1,767 tokens)
    Tool defs:     17 tools & skills  (~1,829 tokens)
     TOTAL: ~3,596 tokens per LLM call

  "3,596 tokens?! That's like flushing doubloons down the drain!
   Let me get me claws on this..."

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  LEFT CLAW — Session History Compression
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Before:  1,826 tokens  (Vercel, auth bugs, Sentry, Slack, Spotify...)
  After:   409 tokens     (just the GitHub-relevant bits)
  Saved:   1,417 tokens (77.6%)
  █████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

  Agagagaga! Now THAT'S what I call a bargain!

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  RIGHT CLAW — Tool & Skill Definition Compression
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Before:  1,743 tokens  (browser, spotify, slack, calendar, email...)
  After:   318 tokens     (github_list_issues + schemas intact)
  Saved:   1,425 tokens (81.8%)
  ███████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

  *eyes pop out* THAT'S BEAUTIFUL! More savings than me secret formula!

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  (V)(;,,;)(V)  MR. KRAB'S FINAL TALLY
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Total before:  3,569 tokens
  Total after:   727 tokens
  Total saved:   2,842 tokens (79.6%)
  ████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

  Money saved at GPT-4o rates ($2.50/1M input tokens):
    Per call:         ~$0.0071
    Per 1,000 calls:  ~$7.11
    Per 100k calls:   ~$710.50

  "80% savings! That's even better than me employee discount
   at the Krusty Krab! Agagagagaga!"

Mr. Krab watched a 15-message conversation about deploying Next.js, fixing auth bugs, setting up CI/CD, configuring Sentry, and playing Spotify — then the user asked about GitHub issues. His left claw kept the GitHub context and snipped away everything else. His right claw kept github_list_issues schemas intact and compressed away 16 irrelevant tool definitions. 80% savings. Every call.

---

The Problem Mr. Krab Solves

Every time OpenClaw sends a message to an LLM, it packs in:

• Full conversation history — every message across WhatsApp, Slack, Discord, Telegram
• 52+ tool/skill definitions — SOUL.md, TOOLS.md, SKILL.md, AGENTS.md, all fully expanded
• System prompts — personality, rules, knowledge base

On a busy session, that's 8,000-15,000+ tokens per call. Most of it irrelevant to what the user just asked.

"That's like ordering the entire menu when you just want a Krabby Patty!"

How Mr. Krab Fixes It

Before Mr. Krab:   ~12,000 input tokens per call
After Mr. Krab:    ~3,500 input tokens per call
                   ─────────────────────────────
                   ~70% savings, every single call

Mr. Krab's compression is query-aware. He knows what the user is asking right now and only keeps what's relevant:

User asks about...	Mr. Krab keeps	Mr. Krab snips away
"Deploy to production"	Deployment discussions, CI/CD tool docs	Spotify skill, Slack commands, unrelated chats
"Play some music"	Spotify skill schema, music preferences	GitHub Issues docs, deployment history
"That bug from yesterday"	Yesterday's debugging messages	Today's small talk, unrelated tool definitions

---

Quick Start

1. Get your API key from ScaleDown

2. Install & configure

export SCALEDOWN_API_KEY="your-api-key"

Add to your ~/.openclaw/openclaw.json:

{
  "extensions": ["@scaledown/openclaw-extension"]
}

3. That's it. Mr. Krab gets to work on every LLM call. Check your logs:

[mr.krab] Claws ready — session: true, tools: true, rate: auto
[mr.krab:left-claw]  3200 -> 1120 tokens (65.0% saved)
[mr.krab:right-claw] 8500 -> 2890 tokens (66.0% saved)

---

How He Works

User sends message -> OpenClaw Gateway
  |
  |-- before_prompt_build hook fires
  |    |
  |    |-- LEFT CLAW ---------> ScaleDown API -----> compressed history
  |    |   (conversation)                            (query-relevant only)
  |    |
  |    |-- RIGHT CLAW --------> ScaleDown API -----> compressed tools
  |        (52+ tool defs)                           (relevant tools only)
  |        (preserve_keywords)                       (names & schemas intact)
  |
  |    Both claws work in parallel -- minimal latency
  |
  |-- Compressed prompt sent to LLM (40-80% fewer tokens)
  |
  '-- llm_output hook -> Mr. Krab logs the savings

The Two Claws

Left Claw (Session History) pinches your conversation transcript per query. A 30-message cross-platform conversation gets snipped down to just the messages that matter for the current question.

Right Claw (Tool/Skill Definitions) pinches the system prompt containing all tool definitions. OpenClaw loads dozens of skills — browser, GitHub, Spotify, Slack, memory, calendar, and more. Mr. Krab compresses the documentation while keeping tool names, parameter schemas, and critical syntax intact via preserve_keywords. He pinches tokens, not tool names.

---

Configuration

All optional. Mr. Krab works great out of the box.

{
  "extensions": {
    "@scaledown/openclaw-extension": {
      "apiKey": "sk-...",
      "rate": "auto",
      "targetModel": "gpt-4o",
      "enableSessionCompression": true,
      "enableToolCompression": true,
      "minTokenThreshold": 500,
      "timeout": 15000
    }
  }
}

Option	Default	Description
apiKey	$SCALEDOWN_API_KEY	Your ScaleDown API key
rate	"auto"	How aggressively Mr. Krab pinches. "auto" recommended. 0.3 = aggressive, 0.7 = gentle.
targetModel	"gpt-4o"	Target LLM for tokenizer-aware compression
enableSessionCompression	true	Left claw on/off
enableToolCompression	true	Right claw on/off
minTokenThreshold	500	Mr. Krab doesn't bother with small fry below this count
timeout	15000	API timeout in ms

---

Safety & Reliability

Mr. Krab is cheap, not reckless:

• Graceful fallback — If the ScaleDown API is down, slow, or returns an error, Mr. Krab silently falls back to uncompressed context. Your assistant keeps working. He'd rather miss a savings opportunity than break the LLM call.
• Threshold gating — Small contexts below minTokenThreshold aren't worth pinching (even Mr. Krab knows some things cost more to save than they're worth).
• Timeout protection — Configurable timeout (default 15s) with AbortController prevents hanging.
• No data storage — ScaleDown compresses in-flight. Your conversation data is not stored. "I'm cheap, not shady."

---

Development

npm install          # Install dependencies
npm run typecheck    # Type check
npm run build        # Build to dist/
npm run dev          # Watch mode
npm test             # Run all 53 tests
npm run demo         # Watch Mr. Krab in action

Project Structure

openclaw-scaledown-extension/
├── demo.ts                    # Mr. Krab's live demo
├── openclaw.plugin.json       # Plugin manifest & config schema
├── package.json
├── tsconfig.json
├── src/
│   ├── index.ts               # Plugin entry — hook registration
│   ├── scaledown-client.ts    # ScaleDown REST API client
│   ├── session-compressor.ts  # Left claw: conversation compression
│   ├── tool-compressor.ts     # Right claw: tool definition compression
│   ├── types.ts               # TypeScript types & defaults
│   └── utils.ts               # Token estimation, serialization, chunking
└── tests/
    ├── utils.test.ts
    ├── scaledown-client.test.ts
    ├── session-compressor.test.ts
    ├── tool-compressor.test.ts
    └── plugin-integration.test.ts

---

Why Mr. Krab?

Unlike static prompt trimming or naive truncation, Mr. Krab uses query-aware semantic compression powered by ScaleDown. He understands what the user is asking and preserves the context that matters — producing better LLM responses with fewer tokens.

• Not truncation — doesn't chop off old messages blindly
• Not summarization — doesn't lose critical details or tool syntax
• Query-aware — adapts compression to every single message
• Two claws — handles both conversation history AND tool definitions
• Never breaks — graceful fallback on any failure

"Are you gonna pinch tokens, or are you gonna waste 'em like Squidward wastes his talent?"

---

License

MIT

(V)(;,,;)(V)