Nova Agent

A high-performance coding agent built in Rust. Designed to be fast, lightweight, and extensible.

Features

• Blazing fast: 4.4MB static binary, <100ms startup, minimal memory footprint
• 6 built-in tools: read (mmap), write, edit (ropey), glob, grep (ripgrep), bash (async)
• Protocol-driven: Clean NACP (Nova Agent Communication Protocol) layer with NDJSON transport
• LLM integration: Anthropic Claude API adapter (extensible to OpenAI and others)
• Interactive REPL: Colored terminal interface with streaming output
• Permission system: Three modes — ask, auto-read, auto-all
• Context management: Automatic compaction for long conversations
• Sandbox support: Pluggable sandbox abstraction (none, process, landlock, docker)

Quick Start

# Build
cargo build --release

# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Interactive mode
./target/release/nova

# Single prompt
nova --prompt "Read main.rs and fix the bug on line 42"

# With options
nova --model claude-sonnet-4-20250514 --permission auto-all --workdir /my/project

Architecture

┌─────────────────────────────────────────────────┐
│              CLI / REPL (nova-cli)               │
├─────────────────────────────────────────────────┤
│         NACP Protocol (nova-protocol)            │
│      Request/Response │ Events │ Transport       │
├─────────────────────────────────────────────────┤
│            Core Engine (nova-core)               │
│   Agent Loop │ Context Manager │ Permissions     │
├─────────────────────────────────────────────────┤
│          Tools (nova-tools) │ Sandbox            │
│  read│write│edit│glob│grep│bash │ (nova-sandbox) │
├─────────────────────────────────────────────────┤
│          LLM Adapter (nova-core::llm)            │
│              Anthropic │ OpenAI                   │
└─────────────────────────────────────────────────┘

Crates

Crate	Purpose
nova-protocol	Message types, tool definitions, NDJSON transport
nova-core	Agent loop, LLM adapters, context management, permissions
nova-tools	Built-in tool implementations (read, write, edit, glob, grep, bash)
nova-cli	CLI binary with interactive REPL
nova-sandbox	Sandbox abstraction layer

CLI Options

nova [OPTIONS]

Options:
  -p, --prompt <PROMPT>          Run a single prompt (non-interactive)
  -m, --model <MODEL>            LLM model [default: claude-sonnet-4-20250514]
      --permission <MODE>        ask | auto-read | auto-all [default: auto-read]
  -d, --workdir <DIR>            Working directory
  -v, --verbose                  Enable debug logging
  -h, --help                     Print help
  -V, --version                  Print version

REPL Commands

Command	Description
/help	Show available commands
/quit	Exit the REPL
/clear	Clear conversation context
/tokens	Show token usage statistics

Configuration

Optional config file at ~/.nova/config.toml:

[llm]
provider = "anthropic"
api_key = "sk-ant-..."
model = "claude-sonnet-4-20250514"

[performance]
threads = 0        # 0 = auto-detect
cache_size = "1GB"

Testing

cargo test          # Run all 29 tests
cargo test -p nova-cli -- --nocapture  # With output

Performance

Metric	Nova Agent	Typical Python agents
Binary size	4.4 MB	100+ MB (with deps)
Startup time	<100ms	3-10s
Memory idle	~5 MB	200-500 MB
Tool call latency	<10ms	50-100ms

License

MIT