MiniMax Code CLI Operational

MiniMax Code CLI is an open-source command-line AI coding tool released by MiniMax. It is a core component of the MiniMax Code client and is open-sourced under the MIT license. In the FrontierHarness Eval, it achieved a task pass rate of 76.7%, with a median completion time of 4 minutes 33 seconds for successful tasks, outperforming the public baseline. The tool supports one-command installation via curl, can connect to IDEs, and focuses on efficient tool calling, permission handling, and enterprise-grade reliability.

Follow us on WeChat and reply with “开源” to join the AI open-source project community
Code editing and verification: Read files, generate diffs, run shell commands and tests in the terminal, and manage tool execution through permission controls and sandboxing.
Flexible model selection: Use a MiniMax account / Token Plan, or connect custom models via BYOK using OpenAI- or Anthropic-compatible API formats.
Multimodal tools: Built-in web search and media tools (mcode-tools), with support for MCP and hosted connectors to handle search and media processing beyond code.
Session continuity and task planning: Resume previous sessions (--continue/--session), use Plan Mode, and advance tasks in parallel with subagents.
Plugin and skill extensions: Extend Agent capabilities through official, local, or GitHub plugins and built-in skills.
Three ways to use it: Interactive TUI (mcode [prompt]), headless mode (mcode exec, suitable for scripts/CI/batch evaluations), and ACP protocol (mcode acp, for IDE editor integration).
Project-level guidance: mcode init . generates AGENTS.md to describe expected outcomes, permitted changes, and verification methods for the project.

Install: On macOS / Linux / WSL, run curl -fsSL https://filecdn.minimax.chat/public/install.sh | bash (on Windows, run irm https://filecdn.minimax.chat/public/install.ps1 | iex in PowerShell) to install the CLI and a compatible Node.js environment in one step.
Verify the installation: Reopen the terminal and run mcode --version to confirm the version. Use mcode --help to view all commands.
Log in or configure credentials: For domestic accounts, run mcode login. For overseas accounts, run mcode login --region global. Complete the login in your browser, then use /status to check the account and /provider to select a model.
Connect your own model (BYOK): Without logging in to MiniMax, set the MCODE_PROVIDER_API_KEY environment variable, then run mcode provider add --name my-provider --base-url ... --api-key-env MCODE_PROVIDER_API_KEY --use to add a custom provider.
Start and submit a task: Enter the project directory with cd /path/to/your/project, run mcode to open the interactive interface and describe the task, or submit it with a one-line command: mcode "Find a failing test, fix the implementation, and run the relevant tests."
(Optional) Generate project guidance: Run mcode init . to generate or update AGENTS.md with the expected task outcome, permitted changes, and verification methods.
Interactive controls: Enter sends a message, @ references files, Shift+Tab toggles Plan Mode, Alt+M toggles permission mode, Esc interrupts a task, and /sessions retrieves previous sessions.
Advanced usage: Use headless mode, mcode exec [prompt], for scripts/CI; use mcode acp for IDE integration; and resume interrupted work with mcode --continue or --session.

Outstanding performance: In the FrontierHarness Eval, it achieved a task pass rate of 76.7% and a median completion time of only 4 minutes 33 seconds for successful tasks. Both metrics outperform the public baseline. Its official positioning is to “continuously unlock model capabilities through excellent harness design.”
Genuinely MIT-licensed open source: Starting with v0.4.12, all first-party source code is open-sourced under the MIT license. Developers can inspect tool-calling and permission-handling logic, while the community can participate in discovering and fixing issues, making it suitable for building enterprise-grade applications.
No model lock-in: Supports both MiniMax accounts / Token Plan and BYOK connections to custom models compatible with OpenAI or Anthropic API formats, avoiding dependence on a single provider.
Flexible three-mode architecture: The same toolset provides an interactive TUI, headless mode, and ACP protocol, covering scenarios from personal development to production engineering workflows.
Comprehensive permission and security mechanisms: Built-in permission modes and sandbox controls let users review diffs and confirm tool calls individually.

GitHub repository: https://github.com/MiniMax-AI/minimax-code

Daily coding and debugging: Ask the Agent in the terminal to locate failing tests, analyze root causes, modify code, and automatically rerun tests to create a “modify-and-verify” loop.
CI/CD and batch tasks: Script tasks with headless mode, mcode exec, and integrate them into continuous integration pipelines, overnight batch refactoring, or large-scale code evaluations.
Deep IDE integration: Use the ACP protocol to connect to editors such as Zed that support the Agent Client Protocol, bringing Agent capabilities into a familiar IDE environment.
Enterprise security review and compliant development: Use the MIT open-source codebase to review tool-calling and permission-handling logic, and combine sandboxing with permission modes to build auditable enterprise applications within controlled boundaries.