Running MiniMax M2.5 Locally with Claude Code
A quick how-to guide for connecting Claude Code to your local MiniMax M2.5 inference server.
24 Feb 2026
TL;DR: Once you have MiniMax M2.5 running locally (see previous post), here's how to connect Claude Code to it.
If you've got MiniMax M2.5 running via the setup guide, connecting Claude Code is straightforward.
Basic Configuration
Edit your ~/.claude.settings.json:
{
"env": {
"ANTHROPIC_BASE_URL": "http://<your-server>:8080",
"ANTHROPIC_AUTH_TOKEN": "any-placeholder-value",
"ANTHROPIC_MODEL": "MiniMax-M2.5",
"ANTHROPIC_SMALL_FAST_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "MiniMax-M2.5",
"CLAUDE_CODE_SUBAGENT_MODEL": "MiniMax-M2.5",
"API_TIMEOUT_MS": "3000000",
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
}
}
Replace <your-server> with your inference server's address.
With Agent Teams and Hooks
If you use agent teams, skills, or hooks:
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1",
"ANTHROPIC_BASE_URL": "http://<your-server>:8080",
"ANTHROPIC_AUTH_TOKEN": "local",
"ANTHROPIC_MODEL": "MiniMax-M2.5",
"ANTHROPIC_SMALL_FAST_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "MiniMax-M2.5",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "MiniMax-M2.5",
"CLAUDE_CODE_SUBAGENT_MODEL": "MiniMax-M2.5",
"API_TIMEOUT_MS": "3000000",
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
},
"hooks": {
"UserPromptSubmit": [
{
"hooks": [
{
"type": "command",
"command": "$HOME/.claude/hooks/skill-activation-prompt.sh"
}
]
}
],
"PreToolUse": [
{
"matcher": "Edit|MultiEdit|Write",
"hooks": [
{
"type": "command",
"command": "$HOME/.claude/hooks/skill-verification-guard.sh"
}
]
}
],
"PostToolUse": [
{
"matcher": "Edit|MultiEdit|Write",
"hooks": [
{
"type": "command",
"command": "$HOME/.claude/hooks/post-tool-use-tracker.sh"
}
]
}
]
},
"skipDangerousModePermissionPrompt": true
}
Next Steps
- Start Claude Code with your configuration
- If prompted, log out and log back in to clear any warnings
- You're running entirely on local hardware
This gives you a local, private inference endpoint that works with Claude Code - useful for analyzing sensitive codebases without cloud dependencies.
Resources
Your next read
Running MiniMax M2.5 Locally on NVIDIA DGX Spark
How I ran a 230B parameter open model on desktop hardware with NVIDIA DGX Spark, Unsloth quantization, and llama.cpp at cloud API speed.
Michael Mueller
Subscribe to Our Monthly AI Native Newsletter
Lessons from our client work, engineering deep-dives, and the AI Native research worth reading.
Monthly. No spam, unsubscribe anytime.
10mins