High-Performance, Lean-Token Senior AI Coding Agent for the Terminal
Quick Start · Key Features · BYOK & Pricing · Slash Commands · 简体中文文档
English · 简体中文
Minimax-coder is an ultra-fast, minimal, and token-frugal terminal coding agent engineered for developers who demand peak coding capability without burning unnecessary tokens or waiting on sluggish cloud round-trips.
Featuring the signature MiniMax pink-to-orange gradient aesthetic (#FF007A → #FF7A00) and high-fidelity waveform terminal branding, Minimax-coder runs as a single standalone binary (./Minimax-coder) or node CLI. It strips away cloud telemetry, daily check-ins, and marketing bloat—leaving a blazing-fast, laser-focused senior engineering agent with universal BYOK support (OpenAI, Google Gemini, Anthropic Claude, DeepSeek, and Ollama) and real-time Chinese Yuan (¥ CNY) cost tracking.
- Inline Tool Output Cap: Restricts tool outputs to 8KB / 200 lines inline; large logs and payloads are seamlessly spilled to
.mcode/artifacts/with concise summaries. - Quiet Test Backpressure: Filters noisy stack traces and duplicate test passes into compact receipts (
PASS (4/4 tests)). - Read Deduplication: Re-reading identical files in the same turn returns an instant memory pointer without consuming prompt tokens.
- Observation Masking: Automatically rolls older tool execution logs into concise markers while preserving conversation logic.
- Terse 7-Tool Core: Exactly 7 minimal schemas (
bash,read_file,edit_file,write_file,glob,grep,git). System prompt + tools use just 722 tokens (under half of the 1,500 token budget).
- Production-Ready Quality: Writes clean, type-safe, modular, performant, and well-architected code following modern software design patterns.
- Mandatory Simplified Chinese Docstrings & Comments: All code comments, inline documentation, and function docstrings are strictly written in Simplified Chinese (简体中文), while preserving standard English identifiers.
- Massive Context Support: Automatically recognizes and unlocks the full context ceilings of your models:
- Google Gemini (
gemini-3.8-flash-high,gemini-3-flash,gemini-2.5-flash): 1,000,000 tokens (1.0M) - Gemini Pro (
gemini-3-pro,gemini-2.5-pro,gemini-1.5-pro): 2,000,000 tokens (2.0M) - Anthropic Claude (
claude-3-7-sonnet,claude-3-5-sonnet): 200,000 tokens (200k) - DeepSeek & OpenAI (
deepseek-v3,deepseek-r1,gpt-4o,o1,o3-mini): 128,000+ tokens (131k) - Default Floor: 131,072 tokens (128k) baseline.
- Google Gemini (
- Universal BYOK: Connect any OpenAI-compatible, Google Gemini, Anthropic Claude, or local Ollama endpoint without vendor lock-in.
- Live Status Bar: Always visible at the bottom of the screen:
🧠 52k/1.0M (5%) · 💰 ¥0.027 · 🔁 3 - Fuzzy Model Matcher: Reads official rates from
pricing.jsonwith support for cache-hit and cache-miss pricing. - Interactive
/costCommand: Displays detailed per-session cost and token breakdowns.
- Single-line tool events with emojis:
✏️ packages/tui/src/index.ts (+23)(-3) ✅ $ pnpm test 2.1s 📖 src/api.ts 120L 🔍 grep "renderWaveform" 4 matches - File additions in green
(+23)and deletions in red(-3), guaranteed to appear on every edit or file write.
- Zero cloud tracking, analytics, or telemetry.
- No daily check-in popups or cloud deploy nudges.
- Standalone self-contained executable: no global node_modules needed.
Launch the pre-compiled native binary directly:
./Minimax-coder# 1. Install dependencies
pnpm install
# 2. Build packages and binary
pnpm build
pnpm run build:binary
# 3. Launch Minimax-coder
./Minimax-coderYou can configure any API endpoint using the CLI:
export GEMINI_API_KEY="your-gemini-key"
./Minimax-coder provider add \
--name gemini \
--base-url https://generativelanguage.googleapis.com/v1beta/openai/ \
--api-format openai-completions \
--model gemini-3.8-flash-high \
--api-key-env GEMINI_API_KEY \
--useexport DEEPSEEK_API_KEY="your-deepseek-key"
./Minimax-coder provider add \
--name deepseek \
--base-url https://api.deepseek.com/v1 \
--api-format openai-completions \
--model deepseek-v3 \
--api-key-env DEEPSEEK_API_KEY \
--useexport ANTHROPIC_API_KEY="your-claude-key"
./Minimax-coder provider add \
--name anthropic \
--base-url https://api.anthropic.com \
--api-format anthropic-messages \
--model claude-3-7-sonnet-20250219 \
--api-key-env ANTHROPIC_API_KEY \
--use./Minimax-coder provider add \
--name local-ollama \
--base-url http://localhost:11434/v1 \
--api-format openai-completions \
--model qwen2.5-coder:32b \
--usePricing is stored in pricing.json in Chinese Yuan (¥ CNY) per 1M tokens:
| Model Family | Cache-Hit Prompt | Cache-Miss Prompt | Output Tokens | Context Window |
|---|---|---|---|---|
| Google Gemini 3.8 Flash High | ¥0.136 / 1M | ¥0.544 / 1M | ¥2.175 / 1M | 1,000,000 (1.0M) |
| Google Gemini 3 Flash | ¥0.136 / 1M | ¥0.544 / 1M | ¥2.175 / 1M | 1,000,000 (1.0M) |
| Google Gemini 3 Pro / 2.5 Pro | ¥2.270 / 1M | ¥9.060 / 1M | ¥36.250 / 1M | 2,000,000 (2.0M) |
| DeepSeek V3 | ¥0.500 / 1M | ¥2.000 / 1M | ¥8.000 / 1M | 131,072 (128k) |
| DeepSeek R1 | ¥1.000 / 1M | ¥4.000 / 1M | ¥16.000 / 1M | 131,072 (128k) |
| MiniMax Text 01 / 02 | ¥1.000 / 1M | ¥1.000 / 1M | ¥1.000 / 1M | 131,072 (128k) |
| Claude 3.7 Sonnet | ¥2.720 / 1M | ¥21.750 / 1M | ¥108.750 / 1M | 200,000 (200k) |
| OpenAI GPT-4o | ¥9.060 / 1M | ¥18.130 / 1M | ¥72.500 / 1M | 128,000 (128k) |
| OpenAI o3-mini | ¥4.000 / 1M | ¥8.000 / 1M | ¥32.000 / 1M | 128,000 (128k) |
| OpenAI o1 | ¥54.380 / 1M | ¥108.750 / 1M | ¥435.000 / 1M | 128,000 (128k) |
| Local Ollama | ¥0.000 | ¥0.000 | ¥0.000 | Local |
Inside the terminal UI, use slash commands to inspect and manage your session:
| Command | Description |
|---|---|
/cost |
Display detailed session cost breakdown (cache hits, misses, output tokens, total ¥ CNY) |
/context |
Inspect current context headroom, active memory, and compaction state |
/diff |
Review current working tree code modifications with compact unified diffs |
/undo |
Revert the last file modification made by the agent |
/test |
Run the workspace test suite with quiet backpressure filtering |
/compact |
Trigger manual context compaction to reclaim token headroom |
/plan |
Switch to architectural plan mode before executing edits |
/provider |
Switch or manage active AI models and providers |
minimax-code/
├── bin/
│ └── Minimax-coder # Main standalone binary entrypoint
├── packages/
│ ├── agent-tools/ # 7 lean tool schemas & token diet primitives
│ ├── config/ # Lean agent configuration & defaults
│ ├── local-runtime-v2/ # Core agent runtime & session orchestration
│ └── tui/ # Terminal interface, cost tracker, ANSI gradient
├── pricing.json # Official model pricing in CNY
└── test/ # Verification test suites
- Senior Staff Engineer Standards: Produces robust, clean, modular, and idiomatic code with full error handling.
- Simplified Chinese Annotations: Every code comment, function docstring, and inline explanation must be in Simplified Chinese (简体中文).
- grep-before-read: Fast regex indexing before opening large files.
- diff-only edits: Targeted hunk modifications, avoiding unnecessary full file overwrites.
This project is licensed under the MIT License.