Crate claude_agent 

claude-agent-rs

A lightweight AI coding agent in Rust that runs Claude on your Max subscription for $0.

Quick Start

use claude_agent::{Agent, AgentConfig};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let agent = Agent::new(AgentConfig::default()).await?;
    let response = agent.chat("what is 2+2?").await?;
    println!("{}", response.text());
    Ok(())
}

Two Inference Backends

CLI mode (default) — pipes through claude -p --output-format json, using your Claude Max subscription. Zero per-token cost.

API mode — direct POST to api.anthropic.com/v1/messages. Set ANTHROPIC_API_KEY and AGENT_BACKEND=api.

Features

• Zero API cost — calls claude -p using your Max subscription
• 8 built-in tools — bash, read, write, edit, glob, grep, web_fetch, agent
• SSE streaming — real-time token output for both API and CLI backends
• MCP integration — full JSON-RPC 2.0 client, connects to all your MCP servers
• Session persistence — SQLite-backed save/restore with --resume
• Multi-model routing — auto-route Opus (complex) vs Haiku (simple) with --auto-route
• Sub-agents — spawn parallel claude -p processes with optional git worktree isolation
• Hooks & skills — lifecycle hooks and /command dispatch
• 4.5MB binary — single static binary with SQLite bundled

Architecture

The agent is organized into 11 layers:

Layer	Name	Purpose
1	Crown	Skill registry + /command dispatch
2	Skull	Lifecycle hooks (pre/post tool use)
3	Security Cage	Permission-gated tool access
4	Brain Core	Dual-backend inference (CLI + API)
5	Frontal Lobe	Context engine + CLAUDE.md injection
6	Temporal Lobe	Sub-agent spawner with worktree isolation
7	Auth Gate	Authentication routing
8	Brainstem	8 built-in tools
9	Root System	MCP server registry (JSON-RPC 2.0)
10	Pedestal	HTTP client (reqwest + rustls)
11	Workbench	Feature flags + session metrics

Using as a Library

use claude_agent::{InferenceEngine, InferenceBackend, Message, Role, MessageContent};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let engine = InferenceEngine::new(Some("sk-ant-..."), InferenceBackend::Api);
    let messages = vec![Message {
        role: Role::User,
        content: MessageContent::Text("Hello!".into()),
    }];
    let request = engine.build_request(&messages, None, None, &[], None);

    // Tokens stream to the callback as they arrive
    let response = engine.chat_stream(&request, &mut |text| {
        print!("{}", text);
    }).await?;

    println!("\nTokens used: {}", response.usage.output_tokens);
    Ok(())
}

Re-exports

pub use auth::AuthGate;

pub use context::ContextEngine;

pub use hooks::HookSystem;

pub use inference::InferenceBackend;

pub use inference::InferenceEngine;

pub use mcp::McpRegistry;

pub use permissions::PermissionEngine;

pub use skills::SkillRegistry;

pub use tools::Tool;

pub use tools::ToolRegistry;

pub use types::AgentConfig;

pub use types::AgentStats;

pub use types::CacheControl;

pub use types::CompactionStrategy;

pub use types::ContentBlock;

pub use types::FeatureFlags;

pub use types::HookEvent;

pub use types::HookResult;

pub use types::InferenceRequest;

pub use types::InferenceResponse;

pub use types::Message;

pub use types::MessageContent;

pub use types::Role;

pub use types::SystemBlock;

pub use types::SystemPrompt;

pub use types::ToolCall;

pub use types::ToolDefinition;

pub use types::ToolResult;

pub use types::Usage;

Modules

agents

auth

context

flags

hooks

inference

mcp

memory

permissions

plugins

sessions

skills

task_queue

telemetry

tools

types

Structs

Agent

High-level agent interface for library consumers.

AgentResponse

Response from a single agent turn.

Constants

MAX_AGENT_DEPTH

Maximum depth for sub-agent spawning to prevent recursive bombs.

MAX_FILE_SIZE

Maximum file size (bytes) the read/write tools will accept.

MAX_OUTPUT_TOKENS

Maximum number of tokens the agent will request in a single turn.

MAX_PROMPT_CHARS

Maximum prompt length in characters (UTF-8) before truncation.

MAX_TOOL_CALLS_PER_TURN

Maximum number of tool calls permitted in a single agent loop.

Functions

redact_error

Redact sensitive environment variable values from error messages.

validate_input

Validate an input string is within bounds and contains no null bytes.