A 1.6 MB agentic coding tool — single static Rust binary, zero runtime dependencies

30+ built-in tools · 11 model providers · 5 platforms prebuilt · MIT

npm i -g @angleschina/angles && angles install

Angles Code CLI is a terminal-native agentic coding assistant compiled into a single static Rust binary (~1.6 MB). It needs no Node, no Python, no runtime, and no compiler toolchain — install in one command and start working. The agent drives 30+ built-in angles-* tools to operate files, directories, the terminal, Git, and the web, switching between 11 model providers at will. Since v0.4.1 it emits AI-generated operation plans before executing: Angles lists high-level steps ("read… install… create… edit… run…") so you always know what it is about to do. Config lives locally under ~/.angles/; API keys are never routed through a relay server.

| | | |---|---| | 📦 Binary size | 1.6 MB (musl static / MSVC) | | 🧩 Built-in tools | 30+ (file / dir / terminal / Git / web) | | 🤖 Providers | 11 (OpenAI / Claude / Gemini / DeepSeek / Grok / Qwen / GLM / Kimi …) | | 🖥️ Prebuilt platforms | 5 (Linux ARM64/x64 · macOS ARM64/x64 · Windows x64) | | ⚡ Runtime dependencies | 0 (pure Rust, no Node / no Python / no libc dynamic linking) | | 🔓 License | MIT |

Why a single binary?

Most agentic coding tools ship as Node/Python packages pulling hundreds of transitive dependencies, or as large multi-layer installers. That works — until you need to reach an ARM64 SBC, a rootless container, or run inside iSH on iOS.

Angles compiles to one static binary with zero runtime dependencies. The distribution problem collapses to "download one file ≤ 1.6 MB and chmod +x". This matters on constrained environments where installing a toolchain is impractical:

| Constraint | Conventional toolchain | Angles | |---|---|---| | Runtime / interpreter required | Node or Python | None (native static binary) | | Install footprint | tens of MB + dependency tree | ≤ 1.6 MB | | Toolchain to build from source | often required | optional (prebuilt binaries exist) | | ARM64 SBC / rootless / iSH | often fails | works |

Architecture overview

angles-cli/
├── src/
│   ├── main.rs          # Entry point & command routing
│   ├── cli.rs           # Clap CLI definitions & help
│   ├── config.rs        # Config load/save/display
│   ├── provider.rs      # 11-provider registry with base URLs
│   ├── gateway.rs       # TUI setup wizard (dialoguer)
│   ├── instructions.rs  # System-prompt template rendering (handlebars)
│   ├── api.rs           # API client (OpenAI/Anthropic/Gemini) + streaming + tool loop
│   ├── search.rs        # Web search URL builder
│   ├── server.rs        # Local HTTP gateway (axum) — `angles serve`
│   └── tools.rs         # 30+ angles-* tool implementations + `doctor`
├── instructions.txt     # System-prompt template (13 KB, {{variable}} injection)
├── AGENTS.md            # Agent tool reference
├── providers.toml       # Provider data source
├── gateway-flow.md      # TUI wizard flow spec
├── docs/                # GitHub Pages site (index/docs/tools/faq.html)
├── Cargo.toml
├── Makefile
├── Cross.toml
└── .github/workflows/release.yml

The conversation loop runs in api.rs: an OpenAI / Anthropic Messages / Gemini-native client with streaming, backed by a tool-calling loop that resolves angles-* commands to the corresponding implementations in tools.rs.

Key design decisions

1. Tool-driven agent loop

Every capability Angles exposes is a first-class angles- command. Rather than letting the model improvise shell one-liners, the agent is bound to a curated, deterministic tool set:

• File: angles-createfile/writefile/appendfile, angles-readfile, angles-replace/replaceall, angles-grep, angles-head/tail …
• Terminal: angles-run (foreground), angles-runbg (background w/ PID), angles-kill
• Git: angles-gitinit/gitcommit/gitlog/gitdiff/gitbranch
• Web: angles-fetch, angles-websearch
• Dir / mgmt: angles-ls/tree/cd/pwd, angles-fileinfo, angles-*dir/copy

Safety model: reads require no approval; writes follow the configured approval policy; deletions and dangerous operations always require user confirmation (see AGENTS.md).

2. Multi-provider, one wire protocol

Provider registry (provider.rs + providers.toml) maps each of the 11 providers to its native wire protocol — OpenAI Chat Completions, Anthropic Messages, or Gemini Native — and normalizes the streaming + tool-call path so the rest of the codebase stays provider-agnostic.

3. AI operation plans (v0.4.1+)

Before touching anything, Angles renders a high-level plan from the system prompt and reports it in the terminal: read… install… create… edit… run…. Every step is visible and auditable; the agent does not operate silently.

4. Local HTTP gateway (v0.2+)

angles serve starts an embedded axum server on http://127.0.0.1:8080 — a web console to edit config, switch providers, and test conversations in the browser. REST surface: /health, /api/config, /api/providers, /api/chat.

5. Privacy-first config

Everything lives in ~/.angles/config.json. API keys can be read from a local file or set via the ANGLES_API_KEY environment variable — never transmitted to a relay.

Configuration

Stored at ~/.angles/config.json:

{
  "language": "zh-CN",
  "provider": "glm",
  "base_url": "https://api.siliconflow.cn/v1",
  "wire_api": "chat",
  "model": "zai-org/GLM-5.2",
  "api_key": "",
  "max_tokens": 16384,
  "daily_token_budget": 1000000,
  "agent_persona": "你是一个专业、高效的编码助手。",
  "search_engine": "bing",
  "search_engine_url": "",
  "approval_policy": "untrusted"
}

API key may also be supplied through the ANGLES_API_KEY environment variable.

Installation

All four methods install the same binary:

npm (recommended)

npm i -g @angleschina/angles && angles install

Routes through the npmmirror registry by default for the fastest fetch in CN.

curl — Linux / macOS / WSL2

curl -fsSL https://zsj305.github.io/angles-cli/install.sh | bash

PowerShell — Windows

irm https://zsj305.github.io/angles-cli/install.ps1 | iex

wget

wget -qO- https://zsj305.github.io/angles-cli/install.sh | bash

When a prebuilt binary is available the installer skips the Rust / compiler toolchain entirely and downloads a ~1.6 MB binary in seconds. Works on iSH, Raspberry Pi, Alpine, and rootless environments.

Quick start

# First-time setup (5-step TUI wizard)
angles gateway

# Interactive session (default mode)
angles

# One-shot non-interactive execution
angles exec "write a Python HTTP server"

# Emit an AI operation plan without executing (v0.4.1+)
angles plan "add JWT auth to this Express project, with tests"

# Start the local HTTP gateway (v0.2+) — http://127.0.0.1:8080
angles serve

# Inspect config / diagnose installation / help
angles config
angles doctor
angles help

Supported providers

| Provider | API Host | Protocol | |---|---|---| | OpenAI | api.openai.com | OpenAI Chat Completions | | Claude (Anthropic) | api.anthropic.com | Anthropic Messages API | | Gemini (Google) | generativelanguage.googleapis.com | Gemini Native API | | DeepSeek | api.deepseek.com | OpenAI Chat Completions | | Grok (xAI) | api.x.ai | OpenAI Chat Completions | | MiniMax | api.minimax.chat | OpenAI Chat Completions | | OpenRouter | openrouter.ai | OpenAI Chat Completions | | Qwen (通义千问) | dashscope.aliyuncs.com | OpenAI Chat Completions | | GLM (智谱) | api.siliconflow.cn | OpenAI Chat Completions | | Kimi (Moonshot) | api.moonshot.cn | OpenAI Chat Completions | | Custom | user-defined | OpenAI / Anthropic / Gemini |

Build from source

# Install Rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Clone & build
git clone https://github.com/ZSJ305/angles-cli.git
cd angles-cli
cargo build --release

# Install
cp target/release/angles ~/.local/bin/

Cross-compile

make setup-arm64 && make arm64   # Linux ARM64
make setup-x64 && make x64       # Linux x64
make macos-arm64                  # macOS ARM64 (macOS only)

Prebuilt binaries for all 5 platforms are produced automatically by GitHub Actions on every tag push — see Releases.

License

MIT — see LICENSE.