Configure

Mini-claude reads its configuration from environment variables. Everything has a sensible default, you can ship without setting any of them.

Environment variables

Variable	Default	What it does
`MINI_CLAUDE_URL`	`http://localhost:11434`	Base URL of the inference server. Must expose `/v1/chat/completions`.
`MINI_CLAUDE_MODEL`	`qwen2.5-coder:7b`	Default model. You can switch at runtime with `/model`.
`MINI_CLAUDE_MODE`	`auto`	Starting mode: `auto`, `explain`, `guided` or `review`. See The 4 modes.
`MINI_CLAUDE_WORKSPACE`	current dir	Root for file tools. Any access outside it asks permission.
`MINI_CLAUDE_TEMPERATURE`	`0.7`	Sampling temperature. `0.0` = deterministic, `1.0+` = creative.
`MINI_CLAUDE_SYSTEM`	(empty)	Overrides the mode’s system prompt. Handle with care.

Examples

Point to a remote server on your LAN and start in review mode:

export MINI_CLAUDE_URL=http://192.168.1.42:11434
export MINI_CLAUDE_MODEL=qwen2.5-coder:7b
export MINI_CLAUDE_MODE=review
go run ./cmd/tui

System prompt: let the modes do their job

Each mode already ships a system prompt tuned to its behavior (see The 4 modes). MINI_CLAUDE_SYSTEM replaces that prompt, you lose the mode’s guardrails. Only use it if you know what you’re doing. To give the agent project context without breaking the modes, prefer a project memory file.

Persisting config

Add your exports to ~/.zshrc or ~/.bashrc, or wrap Mini-claude in a small launcher script:

#!/usr/bin/env bash
export MINI_CLAUDE_MODEL=qwen2.5-coder:7b
export MINI_CLAUDE_MODE=explain
exec mini-claude

Permissions & safety Slash commands