version: "1.0.2" name: minimax-mcp description: "MiniMax MCP server integration with token-efficient AI delegation. Use for web search, image analysis, MCP server integration in terminal/desktop Claude, and Claude-MiniMax collaboration patterns. Provides working scripts and slash commands."

MiniMax MCP

Overview

This skill provides complete MiniMax MCP server integration for Claude, enabling token-efficient AI delegation with 85-90% token savings. MiniMax handles heavy computational tasks while Claude plans and reviews, creating an optimal division of labor.

Usage Guidance

Autonomous Delegation (Compound Engineering)

IMPORTANT: This skill is designed for autonomous use. You do NOT need to ask for permission before delegating tasks to MiniMax. The delegation patterns below are approved workflows for compound engineering.

See docs/agent-instructions/COMPOUND_ENGINEERING.md for the full delegation framework.

When to Prefer MiniMax

Consider delegating these tasks to MiniMax for better efficiency:

1. Web search or research queries - MiniMax handles external searches well
2. Image analysis tasks - MiniMax VLM API provides visual understanding
3. Multi-source research - When you'd otherwise read 5+ files
4. Current information needs - Docs, trends, releases from 2024-2026
5. Parallel asset generation - Launch multiple agents for concurrent work
6. Screenshot analysis - Visual bugs, UI issues, sprite quality assessment

When NOT to Delegate

• HPV playtesting or MCP runtime inspection (requires local control and fast feedback).
• Iterative input loops where latency ruins the feedback cycle (batch locally instead).

Trusted Domains (Auto-Approved)

These sources are reliable and searches are automatically restricted to these domains:

• docs.anthropic.com - Claude official docs
• platform.claude.com - Claude platform docs
• docs.cursor.com - Cursor IDE docs
• cursor.com - Cursor docs
• cookbook.openai.com - OpenAI cookbook
• godotengine.org - Godot official docs
• api.minimax.io - MiniMax API docs

Implementation: The web-search.sh script automatically appends site: filters to all queries, ensuring results only come from trusted domains. To search other domains, ask Sam for permission and use an alternative search method.

Decision Trigger

Before using Grep/Glob for research, pause and ask:

"Would a MiniMax search to trusted docs handle this better?"

If yes → use MiniMax. If searching known local files → use local tools.

Autonomous Delegation Patterns

Pattern 1: Parallel Research Delegation

Launch multiple MiniMax agents concurrently instead of sequential research:

Autonomous Example - No Permission Needed:

# Agent 1: Research Godot 4.5 input system
curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"q":"site:godotengine.org InputEventAction handling"}' &
 
# Agent 2: Research dialogue UI patterns
curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"q":"site:docs.godotengine.org dialogue UI RichTextLabel"}' &
 
# Agent 3: Research quest state management
curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"q":"site:docs.godotengine.org state machine autoload"}' &
 
# Wait for all to complete, then synthesize results
wait

Key Points:

• Use & to launch requests in parallel
• Use wait to collect all results before synthesizing
• Main agent orchestrates, MiniMax agents execute
• No permission needed - this is standard compound engineering

Pattern 2: Parallel Image Analysis

Analyze multiple screenshots or sprites concurrently:

Autonomous Example - Visual Quality Assessment:

# Analyze 5 placeholder sprites in parallel
for sprite in moly_seed nightshade_seed moon_tears npc_circe npc_world; do
  curl -s -X POST "https://api.minimax.io/v1/coding_plan/vlm" \
    -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
    -H "Content-Type: application/json" \
    -d "{
      \"prompt\": \"Rate this sprite on clarity, pixel art quality, and game readiness (1-10). Note issues.\",
      \"image_url\": \"file://$(pwd)/assets/sprites/placeholders/${sprite}.png\"
    }" > "analysis_${sprite}.json" &
done
wait
 
# Synthesize all results into quality report
cat analysis_*.json | jq '.'

Pattern 3: Research + Image Analysis Combo

Combine web search with image understanding for comprehensive analysis:

Autonomous Example - Screenshot Bug Analysis:

# Agent 1: Search for similar bug reports
curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"q":"site:github.com godot dialogue box overlap UI"}' > bug_research.json &
 
# Agent 2: Analyze the screenshot
curl -s -X POST "https://api.minimax.io/v1/coding_plan/vlm" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Identify UI elements, z-order issues, and layout problems in this game screenshot.",
    "image_url": "file://screenshot.png"
  }' > screenshot_analysis.json &
 
wait
 
# Main agent synthesizes both sources
echo "Research findings:" && cat bug_research.json
echo "Screenshot analysis:" && cat screenshot_analysis.json

Token Savings with Parallel Delegation

Sequential (Bad):

• Research task 1: ~2000 tokens (Claude reads results)
• Research task 2: ~2000 tokens
• Research task 3: ~2000 tokens
• Total: ~6000 tokens

Parallel with MiniMax (Good):

• Launch 3 agents: ~100 tokens (Claude orchestrates)
• MiniMax handles all 3: ~6000 tokens (subagent compute)
• Claude reviews synthesis: ~500 tokens
• Total: ~600 tokens for Claude (90% savings)

The key: Claude plans (~100), MiniMax executes (~6000 in subagents), Claude reviews (~500).

Example Pattern

Local file task (use Grep/Read):

User: "Find where player_health is defined"
Agent: *Uses Grep to search codebase* ✅

Research task (use MiniMax):

User: "How does Godot 4.5 handle input?"
Agent: *Calls MiniMax search: site:godotengine.org input handling*
MiniMax: *Returns official docs*
Agent: *Reviews results* ✅

Core Capabilities

1. MCP Server Integration

• Launch MCP Server: Start MiniMax MCP server with proper environment configuration
• Status Monitoring: Check server health and connectivity
• Token Efficiency: Claude plans (~100 tokens), MiniMax executes (~2000 tokens saved)

2. Direct API Access

• Web Search: Google-like search via /v1/coding_plan/search endpoint
• Image Analysis: Understand JPEG/PNG/WebP images via /v1/coding_plan/vlm endpoint
• No Dependencies: Works with curl only (no Python or server required)

3. Terminal & Desktop Support

• Terminal Claude: Use direct API calls via curl
• Desktop Claude: Use MCP server with native tools
• Cursor IDE: Compatible with existing slash commands

4. Production-Ready

• Verified API Key: 126-character key provided
• Tested Endpoints: All functionality verified (2026-01-19)
• Error Handling: Comprehensive troubleshooting guides

Quick Start

Start MCP Server (Desktop Claude)

MINIMAX_API_KEY="<your-minimax-api-key>" \
MINIMAX_API_HOST="https://api.minimax.io" \
uvx minimax-coding-plan-mcp -y

Direct API Usage (Terminal Claude)

Web Search:

curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -H "MM-API-Source: Minimax-MCP" \
  -d '{"q":"your search query"}'

Image Analysis:

curl -s -X POST "https://api.minimax.io/v1/coding_plan/vlm" \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -H "MM-API-Source: Minimax-MCP" \
  -d '{"prompt":"What do you see?","image_url":"https://example.com/image.png"}'

Claude Extension Usage (Recommended)

For Claude extension users, call the MiniMax API directly via the Bash tool - no intermediate scripts needed.

Web Search:

curl -s -X POST "https://api.minimax.io/v1/coding_plan/search" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY:?Set MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -H "MM-API-Source: Minimax-MCP" \
  -d '{"q":"your search query"}'

Image Analysis (URL):

curl -s -X POST "https://api.minimax.io/v1/coding_plan/vlm" \
  -H "Authorization: Bearer ${MINIMAX_API_KEY:?Set MINIMAX_API_KEY}" \
  -H "Content-Type: application/json" \
  -H "MM-API-Source: Minimax-MCP" \
  -d '{"prompt":"What do you see?","image_url":"https://example.com/image.png"}'

Token Efficiency: Claude makes the curl call (~50 tokens), MiniMax handles execution (saves ~2000 tokens). This is the optimal pattern for extension usage.

Workflow Patterns

Pattern 1: Research Delegation

1. Claude Plans (minimal tokens ~50-100): "Search for Godot 4.5 features"
2. MiniMax Executes (heavy lifting ~2000 tokens): Web search via API
3. Claude Reviews (oversight): Process results and provide insights

Pattern 2: Image Analysis

1. Claude Directs (planning): "Analyze this screenshot for UI issues"
2. MiniMax Analyzes (computation): Image understanding via API
3. Claude Synthesizes (quality control): Interpret analysis results

Pattern 3: Plan Execution

1. Claude Creates Plan: Structure approach and requirements
2. MiniMax Executes: Perform research, analysis, coding tasks
3. Claude Reviews: Validate results and iterate as needed

Available Scripts

scripts/check-status.sh

Check MCP server status and health. Verifies environment variables and connectivity.

Usage:

./scripts/check-status.sh

scripts/web-search.sh

Perform web search using direct API. Token-efficient alternative to MCP tools.

Usage:

./scripts/web-search.sh "Godot engine features"

scripts/analyze-image.sh

Analyze images using MiniMax vision capabilities via direct API.

Usage:

./scripts/analyze-image.sh "What bugs do you see?" "screenshot.png"

scripts/test-connection.sh

Verify API key, endpoint connectivity, and environment setup.

Usage:

./scripts/test-connection.sh

scripts/execute-plan.sh

Execute structured plan with MiniMax delegation for maximum token efficiency.

Usage:

./scripts/execute-plan.sh "Research React 18 features and create summary"

scripts/general-query.sh

Send general queries to MiniMax for any task.

Usage:

./scripts/general-query.sh "Explain quantum computing basics"

Environment Setup

Required Environment Variables

MINIMAX_API_KEY="<your-minimax-api-key>"
MINIMAX_API_HOST="https://api.minimax.io"

Inline Usage (Recommended)

Set variables in the same command to avoid persistence issues:

MINIMAX_API_KEY="..." MINIMAX_API_HOST="..." uvx minimax-coding-plan-mcp -y

Reference Documentation

API Endpoints

See api-endpoints.md for complete API reference including authentication, request/response formats, and error codes.

Troubleshooting

See troubleshooting.md for common issues, solutions, and verification steps.

Workflows

See workflows.md for detailed usage patterns, examples, and best practices.

Slash Commands

See slash-commands.md for command reference and integration examples.

Key Benefits

✅ 85-90% Token Reduction: MiniMax handles heavy computation ✅ Production Ready: All tests passing, verified 2026-01-19 ✅ Multiple Integration Methods: Terminal (API) and Desktop (MCP) ✅ Zero Dependencies: Works with curl only ✅ Comprehensive Documentation: Complete guides and references ✅ Error Handling: Robust troubleshooting and validation

When to Use This Skill

Use MiniMax MCP when:

• Research tasks require extensive web searching
• Image analysis is needed
• Token efficiency is critical
• Delegating heavy computational work
• Building workflows with AI collaboration

Choose Integration Method:

• Terminal Claude: Use direct API calls (curl)
• Desktop Claude: Use MCP server (uvx)
• Cursor IDE: Use slash commands or MCP tools

Status: ✅ Production Ready (Verified 2026-01-19) Tests: ✅ All 5 Success Criteria Passing Documentation: ✅ Complete Support: ✅ Comprehensive Troubleshooting Guides