MiniMax Code CLI Open Source: Run a Terminal AI Coding Agent for $0 (MIT + BYOK)

MiniMax open-sourced its terminal AI coding tool, MiniMax Code CLI v0.4.12, under the MIT license. That means two things: the tool itself is free, the code is auditable and modifiable; and more importantly it ships native BYOK (Bring Your Own Key) support — you can point the CLI at any OpenAI / Anthropic-compatible API, including ones with a free tier.

Bottom line: MiniMax Code CLI is free (MIT open source), but running it burns tokens. To run it for $0, go BYOK against a free LLM API — Cerebras (1M tokens/day), Groq (1,000 req/day), or GMI Cloud M2.5 (1M TPM free, no card) — and you get a full terminal coding agent that reads repos, edits code, and runs tests, at zero cost.

Free-Quota Snapshot (official data, Sep 2026)

Key note: MiniMax's own quota is not "free on signup" — it bills via Token Plan or API key. The genuinely $0 route is BYOK against a free API. Here's how.

Three install paths (all $0)

1. Official one-line installer (recommended)

macOS / Linux / WSL:

curl -fsSL https://filecdn.minimax.chat/public/install.sh | bash

Windows (PowerShell):

irm https://filecdn.minimax.chat/public/install.ps1 | iex

It installs to ~/.minimax-code (or %USERPROFILE%\.minimax-code on Windows), needs no sudo / admin, and provisions a compatible Node.js runtime on demand. Reopen your terminal and run mcode --version to verify.

2. npm (if you already have Node)

Requires Node.js 22.19+, 24.2+, 25, or 26:

npm install -g @minimax-ai/code@latest --registry=https://registry.npmjs.org/

3. Alpine / musl Linux is not supported by the installer

Use a manual Node install plus npm on those systems.

Sign-in and official quota

In a project dir, run mcode login (mainland China) or mcode login --region global; finish auth in the browser, then check /status.

Official quota runs on two windows: a rolling 5-hour window + a weekly window. /status shows the remaining percentage and reset countdown for each; /usage gives detail. This is the MiniMax account / Token Plan quota — the free Go/Adagio tiers carry no coding quota, so for $0 go BYOK below.

The free play: BYOK against a free API

This is the core. The CLI natively supports openai-completions, openai-responses, and anthropic-messages API formats — plug in any free LLM API and the whole coding agent runs for $0:

export MCODE_PROVIDER_API_KEY="***"
mcode provider add --name free-src --base-url https://api.example.com/v1   --api-format openai-completions --model model-id   --context-limit 32768 --output-limit 4096 --use
mcode

Three $0 free sources (hands-on, see the linked guides):

Point the CLI at any of these and use /provider to switch and /model to filter, and you run a full coding agent for $0. To stagger across vendors, add several providers (e.g., Cerebras for volume by day, Groq for latency-sensitive work overnight).

Three entry points, three use cases

All three share sessions, models, providers, plugins, and permission controls. For CI, headless mcode exec is the cleanest — bake "fix this failing test" into the pipeline and get one-shot results.

Tips (save tokens, work faster)

• Plan Mode before acting: Shift+Tab to toggle, /plan on|off|status|view to manage. Let it plan on complex tasks to avoid burning tokens going the wrong way.
• Permission modes: /permission offers Ask (confirm each) / Auto (auto-classify risk) / Full access (skip checks). Use Auto day-to-day; Full only in workspaces you fully trust.
• Set a task budget: /goal "finish the migration and run tests" budget=50K — K/M suffixes cap spend; a must for free-tier users.
• Side questions while the main task runs: /btw why this implementation? without interrupting the main task.
• @ to reference files/dirs, Ctrl+V to paste images/video into context.
• Plugin marketplace: mcode plugin marketplace list / mcode plugin add name@official for search, MCP, and more.

Eval numbers, with caveats

Official FrontierHarness Eval report: 76.7% task pass rate (23 passed, 4 failed, 3 timed out unscored of 30) with a 4m 33s median success time — both ahead of the public baselines listed. This is a third-party eval; its scoring and task-time windows differ from public leaderboards, so treat the number as a reference, not an official ranking.

Limits and traps

• "Free" is two-layered: the tool (MIT) is permanently $0; but running the model costs tokens. Official execution is billed via Token Plan / API key. For $0, go BYOK against a free API — don't expect free coding quota from signing up.
• Free Go/Adagio tiers have no coding quota (same as Kimi Code Desktop's Go tier). For $0 coding, use Gemini CLI / DeepSeek-TUI / Tongyi Lingma, or this BYOK route.
• Quota drains on rolling windows (5-hour + weekly); free API BYOK instead is capped by that API's own daily limit (Cerebras 1M/day, Groq 1,000/day, GMI per-minute TPM).
• Alpine / musl Linux can't use the one-line installer — manual Node + npm.
• Data dir defaults to ~/.minimax; BYOK custom providers live here, so multiple CLI installs sharing it will cross-contaminate config.

FAQ

Bottom line

Now that MiniMax Code CLI is open source, "terminal AI coding agent" — something that used to mean paying or heavy setup — becomes a zero-cost thing to run: an MIT tool + BYOK against a free API. To get the most value, stagger Cerebras for volume with Groq for latency, and cap spend with /goal budget. The tool is permanently free, and the free API quotas reset daily / weekly / per-minute — a small project runs for essentially nothing.

Related free sources: Cerebras free API, Groq free LPU, GMI Cloud with MiniMax, Free AI coding tools compared, Kimi Code Desktop.

Figures updated 2026-09-23; versions and free quotas may change — check the official site / repo for the latest.