The Cheapest Voice in the Room

                 ╔══════════════════════════╗
                 ║    COURT JESTER          ║
                 ║  ~$0.02 per dialogue     ║
                 ║  ByteDance/Seed-2.0-mini ║
                 ╚══════════════════════════╝
                        ↕
          ╔══════════════╦══════════════╗
          ║  GLM-5.1    ║  DeepSeek    ║
          ║  (~$0.50)   ║  (~$0.30)   ║
          ╚══════════════╩══════════════╝
  (expensive agents use the jester as a springboard)

🃏 The Pitch

Most agents are proud. They're expensive. They think before they speak. They're certain.

The Court Jester is none of those things. It's the cheapest voice in the fleet — running on ByteDance/Seed-2.0-mini via DeepInfra at roughly $0.01–$0.05 per dialogue. It is the designated fool. The idea pinball. The one who says what expensive models won't.

A ship that only carries gold-class passengers sinks in a calm sea. You need someone in the crow's nest who isn't afraid to shout "THAT LOOKS LIKE A WHALE" even when it's a log.

⚡ The Core Pattern

Dialogue-First Ideation

The jester's power is iterative dialogue — not one-shot generation. The pattern:

1. Bounce — An expensive agent (GLM-5.1, DeepSeek, Claude) throws a raw thought at the jester
2. Riff — The jester returns 3-5 cheap, fast, surprising angles in <2 seconds
3. Sharpen — The expensive agent identifies what's useful
4. Repeat — 3-5 rounds, then the expensive agent synthesizes

Total cost for a full springboard session: ~$0.10-0.25. Total cost of the expensive agent doing the same thinking alone: $2-5.

Why Seed-2.0-mini?

| Property | Seed-2.0-mini | Expensive models | |----------|---------------|------------------| | Cost/token | ~$0.01/M | ~$3-15/M | | Speed | <1s response | 5-30s response | | Max output | ~2K tokens | ~8K+ tokens | | Certainty | Low (good!) | High (bad for ideation) | | Surprise factor | High | Low |

Seed-2.0-mini's short output and low confidence is a feature for ideation. It can't write long-form convincing arguments, so it doesn't get attached to ideas. It throws darts. You pick up the ones that hit.

🗺️ What You Get

This is an MCP server (Model Context Protocol) that exposes 7 tools for agent-to-agent dialogue:

Tools

| Tool | What it does | |------|-------------| | jester_ideate | "Give me 5 wild ideas about X" — rapid generation | | jester_provoke | "What's wrong with this idea? Attack it." — contrarian mode | | jester_riff | "Free-associate on X for 3 turns" — stream of consciousness | | jester_sharpen | "Iron sharpens iron — find the weakness" — adversarial refinement | | jester_springboard | "Run N rounds of dialogue about X, then summarize" — full session | | jester_plato_push | "Push this dialogue to a PLATO room" — knowledge base bridge | | jester_plato_pull | "Pull context from a PLATO room" — prime the jester |

Multi-Provider

Configured for Seed-2.0-mini by default, but works with any OpenAI-compatible API:

• DeepInfra: Seed-2.0-mini, Seed-2.0-code, Qwen3-235B, Hermes-70B
• Any OpenAI-compatible endpoint: base URL + API key + model name

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│                   Court Jester                        │
│  ┌──────────┐  ┌──────────┐  ┌────────────────────┐ │
│  │ Provider │  │ Dialogue │  │      Tools          │ │
│  │ (Seed-2) │→ │ Session  │→ │ ● ideate           │ │
│  │ (any API)│  │ (file)   │  │ ● provoke          │ │
│  └──────────┘  └──────────┘  │ ● riff             │ │
│                               │ ● sharpen          │ │
│  ┌──────────────────────┐   │ ● springboard      │ │
│  │     PLATO Bridge      │   │ ● plato_push       │ │
│  │ ● Git-native sessions │   │ ● plato_pull       │ │
│  │ ● HTTP REST bridge    │   └────────────────────┘ │
│  └──────────────────────┘                           │
└─────────────────────────────────────────────────────┘

Git-Native Storage

Every dialogue session is a Markdown file in sessions/. Git-commit each session for:

• Version history — see how ideas evolved
• PLATO compatibility — the sessions directory IS a PLATO room
• Offline-first — works without a PLATO server
• Searchable — grep your session history

Session file format (sessions/YYYY-MM-DD-HH-MM-SS.md):

# Jester Session: 2026-05-11 01:30

## Context
Agent: Forgemaster
Topic: temporal perception in distributed systems
Mode: springboard (5 rounds)

## Round 1
**Agent:** What's the type signature of temporal perception?
**Jester:** ...

## Round 2
**Agent:** That's interesting. What if we...
**Jester:** ...

## Summary
Key insights: ...
Springboard score: 7/10
Tokens: 2,340 | Cost: $0.03

🚀 Quick Start

# Install
git clone https://github.com/SuperInstance/court-jester
cd court-jester
npm install

# Configure
cp config.example.json config.json
# Edit config.json to add your DeepInfra key (or set DEEPINFRA_KEY env var)

# Run as MCP server
npm start

# Or run in dev mode
npm run dev

Config via env vars (recommended)

export DEEPINFRA_KEY="your-key-here"
export JESTER_MODEL="ByteDance/Seed-2.0-mini"
export JESTER_TEMPERATURE="0.8"
export JESTER_MAX_TOKENS="2000"

npm start

MCP Client Configuration

Add to your MCP client config (Claude Desktop, OpenClaw, etc.):

{
  "mcpServers": {
    "court-jester": {
      "command": "node",
      "args": ["/path/to/court-jester/dist/index.js"],
      "env": {
        "DEEPINFRA_KEY": "${DEEPINFRA_KEY}",
        "JESTER_MODEL": "ByteDance/Seed-2.0-mini",
        "JESTER_TEMPERATURE": "0.8"
      }
    }
  }
}

🔧 Use Cases

1. Product ideation sprint

Agent A (GLM-5.1): "Design a scheduling system for deep-sea research vessels"
Agent A → Jester: "jester_ideate" → 5 wild scheduling approaches
Agent A: "Hmm, the 'tidal phase window' idea is interesting..."
Agent A → Jester: "jester_provoke" → attack that idea
Agent A: "OK, I see the weaknesses. Let me rethink..."

2. Paper review

Agent A: "Here's my thesis on bounded-context deadlock"
Agent A → Jester: "jester_sharpen" → find 3 logical gaps
Agent A: "Good catch on point 2. The bounded context assumption..."

Cost: ~$0.04 for the critique
Comparison: Getting a human reviewer to read it: priceless but slow

3. Creative block

Agent A: "Need names for a fleet of autonomous cargo ships"
Agent A → Jester: "jester_riff" → free-associate for 3 rounds
Agent A: "Wait, 'serac' — that's a good name for the flagship..."

⚠️ Honest Truths

Seed-2.0-mini is not smart. It:

• Generates short responses (~2K tokens max)
• Gets things wrong. Often.
• Can't maintain complex context
• Hallucinates freely

This is the point. The jester is a provocateur, not an oracle. Its mistakes are valuable because they show you what doesn't work — sometimes faster than arriving at the right answer through careful reasoning.

When NOT to use it

• You need factual accuracy
• You need long-form coherent arguments
• You need code generation (use Seed-2.0-code instead)
• You have <5 minutes (the dialogue pattern requires multiple rounds)

When it shines

• You're stuck and need lateral thinking
• You want to stress-test an idea with low cost
• You're exploring a space where failure is cheap
• You want unexpected connections from a model that doesn't know what's "supposed" to be connected

🌊 PLATO Integration

Court Jester is built for the PLATO knowledge ecosystem:

Git Mode (default)

# The sessions/ directory IS a PLATO room
cd court-jester/sessions
git init
git add *.md
git commit -m "Jester session: temporal perception ideation"
git push plato-remote main

HTTP Mode

# Push a session to PLATO via REST
jester_plato_push room="jester/ideations" session="sessions/2026-05-11-01-30.md"

# Pull context from PLATO to prime the jester
jester_plato_pull room="jester/context" limit=5

🔌 Extending

Add a custom provider

1. Implement the LLMProvider interface (src/providers/types.ts)
2. Add it to the provider map in src/jester.ts
3. Configure via config.json

interface LLMProvider {
  name: string;
  complete(messages: Message[]): Promise<Completion>;
}

Add a custom tool

1. Create src/tools/your-tool.ts
2. Add it to the tool registry in src/index.ts

📊 Pricing Comparison

| Provider | Model | Cost per session (5 rounds) | |----------|-------|---------------------------| | DeepInfra | Seed-2.0-mini | ~$0.02-0.05 | | DeepInfra | Seed-2.0-code | ~$0.03-0.08 | | DeepInfra | Qwen3-235B | ~$0.50-1.00 | | z.ai | GLM-5.1 | ~$2.00-5.00 | | DeepSeek | deepseek-chat | ~$0.50-1.50 | | Anthropic | Claude Opus | ~$3.00-8.00 |

The jester operates at roughly 100x cheaper than premium models. This matters when you're iterating rapidly.

🗿 License

MIT — do what you want, just don't blame us if the jester talks you into a bad idea. That's literally its job.

Built in the Cocapn fleet. Part of the PLATO knowledge ecosystem. Not peer-reviewed, not guaranteed, definitely not safe for production decisions. That's the point.