MiniMax has released the source code for MiniMax Code CLI v0.4.12 on GitHub, moving its command-line AI programming tool from closed product to open-source. The company frames the move as a step toward enterprise-grade security: developers can now inspect the exact permission checks and tool-call logic that AI agents use when writing code and manipulating files in production environments.
In the FrontierHarness Eval benchmark, the CLI achieved a 76.7% task success rate with a median completion time of 4 minutes 33 seconds for successful tasks, outperforming the public baseline on both metrics. The repository also serves as the official CLI for the MiniMax AI Platform, supporting text, image, video, speech, vision, and web search through a single mmx command, with one-click integration into Claude Code, Cursor, Codex, and other AI agents.
Confirmed
- Repository: github.com/MiniMax-AI/cli (MIT license)
- Version released: v0.4.12
- Benchmark: FrontierHarness Eval — 76.7% task success rate, median 4 min 33 sec per successful task
- Install: npm install -g mmx-cli (requires Node.js 18+)
- Auth: mmx auth login (OAuth with PKCE or API key); auto-detects Global (api.minimax.io) vs CN (api.minimaxi.com) region
- Agent integration: npx skills add MiniMax-AI/cli -y -g for Claude Code, OpenClaw, Cursor, Codex, Grok Build, OpenCode, Hermes, Pi
- Capabilities: text chat (streaming, system prompts, JSON output), image generation (aspect ratio, batch), video generation (Hailuo-2.3 and MiniMax-H3 models, async with task polling), speech synthesis (30+ voices, streaming), speech-to-text (asr-1.0, multiple formats), vision (image understanding), web search
- Token Plan required: subscription on platform.minimax.io (Global) or platform.minimax.cn (CN)
- Config: ~/.mmx/config.json; environment variables for region, base URL, output format, timeout, verbose logging
Unknown
- Exact release date of v0.4.12 (GitHub shows SemVer tag workflow; no calendar date in source)
- FrontierHarness baseline composition and whether the 76.7% figure has been independently replicated
- Pricing tiers and quota limits for Token Plan (docs link to pricing page without inline numbers)
- SLA, uptime guarantees, or enterprise support terms for the CLI
- Roadmap for future model additions beyond current CLI support
Our take
Open-sourcing shifts the trust boundary: enterprises can audit the permission checks and tool-call paths that AI agents execute in their repos. The 76.7% FrontierHarness score is vendor-reported; treat it as directional until independent teams replicate it. The real adoption lever is the one-click agent setup—if it reliably configures Claude Code and Cursor without manual YAML editing, that lowers friction for teams already using those agents.
