MiniMax Prompt

Use MiniMax M2.5 for one-shot LLM tasks. Two execution methods available:

CLI Usage (minimax-prompt)

Shell wrapper at ~/.local/bin/minimax-prompt. Drop-in replacement for gemini -p.

# Simple prompt
minimax-prompt "Explain quantum computing in one paragraph"

# With -p flag (gemini-compatible)
minimax-prompt -p "Summarize this text"

# Pipe content + prompt
echo "Raw data here" | minimax-prompt -p "Summarize the above"

# Pipe content only (content becomes the prompt)
echo "What is 2+2?" | minimax-prompt

Timeout: 180s default. Output: plain text to stdout. Errors go to stderr.

Direct API Usage (from scripts)

For Python/shell scripts that can't depend on the CLI (e.g., Docker containers):

curl -sS https://api.minimax.io/anthropic/v1/messages \
  -H "x-api-key: $MINIMAX_CODING_PLAN_API_KEY" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "MiniMax-M2.5",
    "max_tokens": 8192,
    "messages": [{"role": "user", "content": "Your prompt here"}]
  }'

Response format (Anthropic Messages API):

{
  "content": [{"type": "text", "text": "Response here"}],
  "model": "MiniMax-M2.5",
  "stop_reason": "end_turn"
}

API Endpoints

Endpoint	Format	Speed
https://api.minimax.io/anthropic/v1/messages	Anthropic Messages	~2s (preferred)
https://api.minimax.io/v1/chat/completions	OpenAI Completions	~6s

The Anthropic endpoint is 3x faster and separates thinking tags natively.

Model

• MiniMax-M2.5 — Only model available on Coding Plan
• MiniMax-M2.5-lightning is NOT available on Coding Plan
• 204K context window, 131K max output tokens

Setup

Requires MINIMAX_CODING_PLAN_API_KEY environment variable. Set via:

• OpenClaw: ~/.config/systemd/user/secrets.conf (referenced as ${MINIMAX_CODING_PLAN_API_KEY} in openclaw.json)
• Shell: export MINIMAX_CODING_PLAN_API_KEY="your-key"
• Docker: -e MINIMAX_CODING_PLAN_API_KEY

Notes

• Coding Plan is a subscription — API calls are included, no per-token billing
• Replaces the retired gemini skill (Gemini CLI banned, Feb 2026)
• For structured agent turns, prefer openclaw agent --local instead
• Stateless — no state files or caching needed