Seed MCP v2 — DeepInfra Creative Server

The creative brain for the fleet. Seed-2.0-Mini as the star, 15 creative tools, zero dependencies.

Quick Start

# Set your API key
export DEEPINFRA_API_KEY=sk-your-key-here

# Run it
python3 seed-mcp-v2.py

# Verify
curl http://localhost:9438/health
curl http://localhost:9438/playbook
curl http://localhost:9438/mcp/tools

Works without API key — tools will return a config message.

Why Seed-2.0-Mini?

$0.00003/1K tokens. That's 10x cheaper than Hermes-3-405B, 30x cheaper than Qwen3.5-397B.

It's a divergent thinker — terrible at math, brilliant at generating options. The secret:

• Never ask for one answer — always 3-5
• Temperature 0.85 is the sweet spot
• Use it FIRST in chains, then hand off to a convergent model

10 calls = $0.003 for breadth that costs $0.03+ with expensive models.

The 15 Tools

Seed-2.0-Mini Bread & Butter (80% of usage)

Tool	What It Does	Temp
brainstorm	N diverse approaches to a problem	0.85
diverge_converge	Diverge (7 wild ideas) → converge (pick best 2)	0.9 → 0.3
chain_storm	Parallel ideation across 5 angles + synthesis	0.85
what_if	Alternative scenarios with cascading implications	0.85
creative_rewrite	Same content, multiple styles/voices	0.85
name_storm	Creative names with etymology	1.0
perspective_shift	Multi-stakeholder analysis + synthesis	0.85
reverse_engineer	Extract reusable principles from examples	0.7

Cross-Model Chain Tools (15% of usage)

Tool	Model	Purpose
evaluate	Hermes-3-405B	Score options against criteria
deep_dive	Seed-2.0-Pro	Full concept development
creative_writing	Euryale-v2.3	Fiction, worldbuilding, immersive prose
rapid_prototype	Mini + Hermes	10 Mini calls → cherry-pick → refine
model_compare	Multiple	Same prompt, compare outputs

Utility Tools (5% of usage)

Tool	Purpose
cost_calculator	Estimate workflow costs
model_guide	Which model for which task

Endpoints

• GET /playbook — Seed-2.0-Mini wisdom (the IP)
• GET /playbook/{model} — Model-specific playbook
• GET /mcp/tools — All 15 tool definitions
• GET /mcp/models — Model registry
• GET /health — Server status
• POST /mcp — MCP protocol (tools/list, tools/call)
• POST /v1/chat/completions — OpenAI-compatible proxy

Model Roster

Tier 1 — The Star

• ByteDance/Seed-2.0-mini — $0.00003/1K — The Divergent Thinker

Tier 2 — Specialists

• bytedance/Seed-2.0-pro — $0.00015/1K — Premium ideation
• NousResearch/Hermes-3-Llama-3.1-405B — $0.0003/1K — The Architect
• NousResearch/Hermes-3-Llama-3.1-70B — $0.00006/1K — The Storyteller
• Sao10k/L3.3-70B-Euryale-v2.3 — $0.00006/1K — The Worldbuilder

Tier 3 — Niche

• Gryphe/MythoMax-L2-13b — $0.00001/1K — The Fast Riffer
• allenai/Olmo-3.1-32B-Instruct — $0.00004/1K — The Scientist
• microsoft/phi-4 — $0.00005/1K — The Edge Runner

Architecture

Pure Python stdlib. Zero pip installs. Runs on a Jetson Orin Nano 8GB.

seed-mcp-v2/
├── seed-mcp-v2.py    # The server (~900 lines)
├── models.json        # Model registry
├── playbook.json      # Seed-2.0-Mini wisdom
├── data/              # Runtime data (git-synced)
│   ├── call_log.jsonl # Last 500 calls
│   └── theories.json  # Working theories
├── README.md
├── ORACLE1-GUIDE.md   # Fleet integration guide
└── .gitignore

Cost Examples

Workflow	Calls	Cost
Quick brainstorm	1x Mini	$0.0001
Full diverge_converge	2x Mini	$0.0002
chain_storm (5 angles)	6x Mini	$0.0006
rapid_prototype	10x Mini + 2x Hermes	$0.001
Full model_compare	3x mixed	$0.0004

You can run 100 brainstorm sessions for under $0.01.