claude-agent-rs

A lightweight Rust binary that runs Claude as a full coding agent. Uses your Max subscription for $0 API costs, or pay-per-token via the API.

Install

cargo install claude-agent-rs

Or build from source:

git clone https://github.com/ExpertVagabond/claude-agent-rs.git

cd claude-agent-rs

cargo build --release

./target/release/claude-agent

Requirements

Rust 2021 edition (for building)
One of:
- claude CLI installed + Claude Max subscription (CLI backend, $0)
- ANTHROPIC_API_KEY + AGENT_BACKEND=api (API backend, pay-per-token)

Usage

Interactive REPL

$ claude-agent
  claude-agent v0.8.0 (Rust)
  claude-opus-4-6 · backend: cli (Claude Max)
  8 tools · no MCP servers

❯ write a fizzbuzz in Rust, compile it, run it
  ▸ write: /tmp/fizzbuzz.rs
  ✓ File written
  ▸ bash: rustc /tmp/fizzbuzz.rs -o /tmp/fizzbuzz && /tmp/fizzbuzz
  ✓ 1, 2, Fizz, 4, Buzz...

One-shot mode

claude-agent -p "list all TODO comments in this project"

CLI options

claude-agent                     Start interactive REPL
claude-agent -p "prompt"         One-shot mode (run and exit)
claude-agent [directory]         Start REPL in directory
claude-agent --resume <id>       Resume a previous session
claude-agent --auto-route        Auto-select Opus vs Haiku by complexity
claude-agent --system-prompt "…" Override the system prompt
claude-agent --help              Show help
claude-agent --version           Show version

Built-in Tools

Tool	Description
bash	Execute shell commands with timeout
read	Read files with line numbers
write	Create or overwrite files
edit	Exact string replacement
glob	Pattern-based file search
grep	Regex content search (uses ripgrep)
web_fetch	HTTP GET requests
agent	Spawn sub-agents with optional git worktree isolation

Two Inference Backends

CLI mode (default) — Pipes prompts to claude -p --output-format stream-json. Uses your Claude Max subscription's OAuth token. Zero per-token billing.

API mode — Direct POST to api.anthropic.com/v1/messages with SSE streaming. Set AGENT_BACKEND=api and ANTHROPIC_API_KEY.

Features

8 built-in tools — bash, read, write, edit, glob, grep, web_fetch, agent
SSE streaming — real-time token output for both backends
MCP client — full JSON-RPC 2.0 stdio transport, auto-discovers tools from ~/.mcp.json
Session persistence — SQLite-backed save/restore with --resume
Sub-agents — spawn parallel claude -p processes with optional git worktree isolation
Multi-model routing — auto-select Opus (complex) vs Haiku (simple) with --auto-route
Skills — user-defined /commands loaded from ~/.claude-agent/skills/*.md
Plugins — custom tools via ~/.claude-agent/plugins/ (manifest.json + executable)
CLAUDE.md injection — reads project + home CLAUDE.md into system prompt
Context management — auto-compaction when approaching token limits
Permissions — blocks dangerous commands (rm -rf /, fork bombs, etc.)
Library API — use as a crate: Agent::new(config).await?.chat("...").await?

Configuration

MCP Servers

Place your MCP config at ~/.mcp.json:

{
  "mcpServers": {
    "my-server": {
      "command": "node",
      "args": ["path/to/server.js"]
    }
  }
}

Custom Skills

Create .md files in ~/.claude-agent/skills/:

# Review code for security issues

Review the current codebase for security vulnerabilities. Focus on: {args}

Then use /review OWASP top 10 in the REPL.

Custom System Prompt

Place a system.md at ~/.claude-agent/system.md to override the default agent identity.

Stats

Binary: ~5MB (release, LTO, stripped, SQLite + rustls bundled)
Dependencies: 14 crates
Startup: <100ms
Zero warnings — clean clippy + rustc

License

MIT

claude-agent-rs 1.0.0