kaze

A memory-minimal AI coding agent for the terminal. Written in Rust. MIT licensed. Early development.

The name means "wind" in Japanese.

Why kaze

Most AI coding tools carry serious memory overhead.

Tool	Idle	Peak
Claude Code (Node.js)	~300 MB	8.5 GB+ (known leaks)
OpenCode (TypeScript)	40-80 MB	~400 MB
kaze (Rust)	<25 MB	<80 MB

kaze is built on rig-core for LLM abstraction and runs on a single-threaded tokio runtime. The release binary is optimized for size with opt-level="z", LTO, and symbol stripping.

Quick Start

# Install
cargo install --path .

# Set an API key (pick your provider)
export ANTHROPIC_API_KEY="your-key-here"

# One-shot question
kaze ask "explain ownership in rust"

# Interactive chat
kaze chat

# Use a different provider or model
kaze ask --provider openai "hello"
kaze ask --model ollama/llama3 "hello"

# List available models
kaze models

Features

CLI

• kaze ask "question" for one-shot streaming responses
• kaze chat for interactive multi-turn REPL
• kaze session list, kaze session resume, kaze session delete, kaze session new
• kaze config show to view current configuration
• kaze models to list available models per provider (marks the default)
• --provider and --model flags on ask and chat
• provider/model shorthand (e.g., openai/gpt-4.1, ollama/llama3)

Providers

• Anthropic (default)
• OpenAI
• OpenRouter (supports many models, including free tiers)
• Ollama (local, no API key needed)

Context Management

• Token counting with BPE tokenization (tiktoken-rs), displayed after each response
• Per-model context window limits with warning at 80% and auto-truncation at 95%
• Context compaction: LLM-based summarization of old messages, triggered manually (/compact) or automatically at 90% usage

Tools and Agent Loop

• Built-in tools: read_file, write_file, edit, glob, grep, bash
• read_file: path validation, size limits, binary detection
• write_file: full-file writes with parent directory creation
• edit: search-and-replace with exact text matching and diff output
• glob: pattern matching, contained to project root
• grep: regex content search with file filtering and match limits
• bash: shell execution with timeout, output cap, and env variable filtering
• Tool framework: Tool trait and ToolRegistry with JSON Schema definitions for LLM function calling
• Agent loop: kaze ask autonomously calls tools in a multi-turn cycle, executing tool calls and feeding results back until the LLM produces a final answer

Sessions

• Conversations saved as JSONL files at ~/.local/share/kaze/sessions/
• Sessions survive restarts
• Resume by full or partial ID (git-style short IDs)
• kaze chat --session {id} to resume directly

Configuration

• TOML config at ~/.config/kaze/config.toml (XDG paths)
• Per-project override via kaze.toml in project root
• Environment variable resolution with {env:VAR} syntax
• Configurable system prompt via system_prompt

Chat

• Readline support: arrow keys, history recall, Ctrl+R search
• Persistent readline history across sessions
• Slash commands: /history, /clear, /compact, /help
• Markdown-lite formatting for responses (bold, inline code, fenced code blocks)
• Streaming token-by-token output

Permissions and Safety

• Per-tool permissions: allow, ask, or deny via [permissions] in config
• Interactive prompts for sensitive tools (bash defaults to ask)
• Session-level "always allow" option
• Wildcard matching for bash commands (e.g., allow git *, deny rm *)
• Colored unified diffs shown before file writes and edits, with confirm/reject when permission is ask

Configuration

Global config at ~/.config/kaze/config.toml. Per-project override with kaze.toml in your project root.

model = "claude-sonnet-4-6"
system_prompt = "You are a senior Rust developer. Be concise and precise."

[provider.anthropic]
api_key = "{env:ANTHROPIC_API_KEY}"

[provider.openai]
api_key = "{env:OPENAI_API_KEY}"

[provider.openrouter]
api_key = "{env:OPENROUTER_API_KEY}"

[provider.ollama]
base_url = "http://localhost:11434"

[compaction]
auto = true
keep_recent = 4
reserved = 10000

[permissions.tools]
read_file = "allow"
glob = "allow"
grep = "allow"
write_file = "allow"
edit = "allow"
bash = "ask"

[permissions.bash_commands]
"git status" = "allow"
"cargo build" = "allow"
"rm *" = "deny"

Roadmap

Built incrementally across 8 phases.

Phase	Description	Status
0	Project scaffold	Done
1	Core (ask, streaming, config)	Done
2	Multi-turn chat + sessions	Done
3	Multi-provider (OpenAI, OpenRouter, Ollama)	Done
4	Context management (token counting, compaction)	Done
5	Tools (read, write, edit, grep, bash)	Done
6	Agent loop	Done
7	TUI (ratatui)	In Progress
8	Advanced (MCP, custom agents, rules)	Planned

Known Limitations

• TUI sidebar is empty (no session navigation yet)
• TUI trust level defaults to Safe (bash tool disabled in TUI mode)
• config set uses toml::Value internally (no comment preservation in config files)

Inspired by OpenCode and aichat.

License

MIT