Claude Agent SDK for Rust

Rust bindings for the Claude Agent SDK protocol used by the Claude Code CLI.

This crate provides a Tokio-native SDK for driving an authenticated local claude CLI process with typed options, message parsing, interactive sessions, control requests, SDK MCP servers, and local/session-store helpers. It is modeled after the Python claude-agent-sdk API while using Rust-native async, traits, and type-safe data structures.

Status

This repository is an SDK implementation, not an official Anthropic package. It targets parity with the Python Claude Agent SDK public behavior and is currently verified by the local test suite.

Verified before publication:

cargo test
cargo check --features otel

The normal test suite avoids paid/authenticated CLI calls. The ignored e2e_cli smoke test can be run separately when the Claude CLI is authenticated.

Requirements

• Rust 1.70+
• Tokio runtime
• Claude Code CLI installed and authenticated
• claude available on PATH, or set via ClaudeAgentOptions::builder().cli_path(...)

Installation

Use the Git repository directly:

[dependencies]
claude-agent-sdk = { git = "https://github.com/PandelisZ/claude-agent-sdk-rust" }

Optional OpenTelemetry propagation support:

[dependencies]
claude-agent-sdk = { git = "https://github.com/PandelisZ/claude-agent-sdk-rust", features = ["otel"] }

Quick start

One-shot query

use claude_agent_sdk::query;

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let result = query("Explain ownership in Rust in two sentences.", None).await?;
    println!("{}", result.content);
    Ok(())
}

Full message sequence

use claude_agent_sdk::query_messages;

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let messages = query_messages("List the files you would inspect first.", None).await?;

    for message in messages {
        println!("{message:?}");
    }

    Ok(())
}

Streamed one-shot input

use claude_agent_sdk::query_stream_messages;

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let stream = futures::stream::iter(vec![
        serde_json::json!({
            "type": "user",
            "message": {"role": "user", "content": "First prompt frame"}
        }),
        serde_json::json!({
            "type": "user",
            "message": {"role": "user", "content": "Second prompt frame"}
        }),
    ]);

    let messages = query_stream_messages(stream, None).await?;
    println!("received {} messages", messages.len());
    Ok(())
}

Interactive client

The interactive client mirrors the Python SDK's explicit connection model: create the client, call connect(), send prompts or control requests, receive messages, then disconnect() or close().

use claude_agent_sdk::{ClaudeAgentClient, ClaudeAgentOptions, PermissionMode};

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let options = ClaudeAgentOptions::builder()
        .model("claude-sonnet-4-20250514")
        .permission_mode(PermissionMode::AcceptEdits)
        .build();

    let mut client = ClaudeAgentClient::new(options)?;
    client.connect().await?;

    client.query("Inspect this repository and summarize the test strategy.").await?;
    let messages = client.receive_response().await?;

    for message in messages {
        println!("{message:?}");
    }

    client.disconnect().await?;
    Ok(())
}

Convenience connection helpers are also available:

client.connect_with_prompt("Start by reading Cargo.toml").await?;
client.connect_with_stream(prompt_stream).await?;

Streaming responses

use claude_agent_sdk::{ClaudeAgentClient, ClaudeAgentOptions, StreamEvent};

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let mut client = ClaudeAgentClient::new(ClaudeAgentOptions::default())?;
    client.connect().await?;

    let mut events = client.stream_message("Write a short haiku about compilers.").await?;
    while let Some(event) = events.recv().await {
        match event {
            StreamEvent::ContentChunk(text) => print!("{text}"),
            StreamEvent::Complete(response) => println!("\nfinished: {:?}", response.stop_reason),
            StreamEvent::Error(error) => eprintln!("error: {error}"),
            _ => {}
        }
    }

    client.disconnect().await?;
    Ok(())
}

Configuration

Use ClaudeAgentOptions::builder() for ergonomic configuration.

use claude_agent_sdk::{
    ClaudeAgentOptions, MCPServerConfig, PermissionMode, SettingSource, SkillsConfig,
};
use std::collections::HashMap;

let mut env = HashMap::new();
env.insert("RUST_LOG".to_string(), "info".to_string());

let options = ClaudeAgentOptions::builder()
    .cwd("/path/to/project")
    .model("claude-sonnet-4-20250514")
    .system_prompt("You are a careful Rust coding agent.")
    .allowed_tools(vec!["Read".into(), "Edit".into(), "Bash".into()])
    .permission_mode(PermissionMode::AskOnFirstUse)
    .setting_sources(vec![SettingSource::User, SettingSource::Project])
    .skills(SkillsConfig::All)
    .env(env)
    .mcp_server(
        "docs",
        MCPServerConfig::Stdio {
            command: "docs-mcp".into(),
            args: Some(vec!["--stdio".into()]),
            env: None,
        },
    )
    .build();

Supported option areas include:

• CLI path, working directory, environment, extra CLI args
• model, fallback model, betas, thinking config, task budgets
• system prompt presets/files, dynamic prompt controls, skills config
• allowed/disallowed tools and permission mode
• can_use_tool callback support for streaming/client mode
• hooks and hook event inclusion
• stdio, SSE, HTTP, raw/path MCP server config
• in-process SDK MCP servers
• sandbox settings
• session resume, continue, fork, and session-store mirroring
• stderr callback
• optional OpenTelemetry context propagation with the otel feature

SDK MCP servers

The crate can host in-process SDK MCP tools and bridge them through Claude Code control requests.

use claude_agent_sdk::{create_sdk_mcp_server, tool, ClaudeAgentOptions, MCPContent};

let server = create_sdk_mcp_server(
    "local-tools",
    vec![tool(
        "greet",
        "Greet a user",
        serde_json::json!({
            "type": "object",
            "properties": {"name": {"type": "string"}},
            "required": ["name"]
        }),
        |input| {
            let name = input.get("name").and_then(|v| v.as_str()).unwrap_or("there");
            Ok(vec![MCPContent::Text { text: format!("Hello, {name}!") }])
        },
    )],
);

let options = ClaudeAgentOptions::builder()
    .sdk_mcp_server("local-tools", server)
    .build();

Use create_sdk_mcp_server_with_version and tool_with_annotations when the CLI needs server version metadata or MCP tool annotations.

Permission callbacks

use claude_agent_sdk::{ClaudeAgentOptions, PermissionResult};

let options = ClaudeAgentOptions::builder()
    .can_use_tool(|tool_name, input, _context| async move {
        if tool_name == "Bash" {
            return Ok(PermissionResult::deny("Bash is disabled in this run"));
        }
        println!("allowing {tool_name} with input {input:?}");
        Ok(PermissionResult::allow())
    })
    .build();

can_use_tool requires streaming/client mode. String-based one-shot query_messages(...) rejects this callback because there is no live bidirectional control channel after spawning the subprocess.

Hooks

Hooks are represented as callback matchers grouped by event name.

use claude_agent_sdk::{ClaudeAgentOptions, HookCallback, HookMatcher};

let callback = HookCallback::new(|input, _tool_use_id, _context| async move {
    println!("hook input: {input:?}");
    Ok(serde_json::json!({"continue": true}))
});

let options = ClaudeAgentOptions::builder()
    .hook("PreToolUse", HookMatcher::new(callback).matcher("Edit"))
    .include_hook_events(true)
    .build();

Sessions

Local Claude transcript helpers are available from the crate root and from claude_agent_sdk::sessions.

use claude_agent_sdk::{get_session_info, get_session_messages, list_sessions};

#[tokio::main]
async fn main() -> claude_agent_sdk::Result<()> {
    let sessions = list_sessions(None).await?;

    for session in sessions {
        println!("{} {:?}", session.id, session.title);
    }

    let info = get_session_info("session-uuid", None).await?;
    let messages = get_session_messages("session-uuid", None).await?;

    println!("{info:?}");
    println!("{} messages", messages.len());
    Ok(())
}

Session helpers include:

• list, inspect, page, and delete local sessions
• list and read subagent transcripts
• rename and tag sessions by appending metadata events
• fork sessions with remapped IDs
• import local transcripts into a custom session store
• store-backed list/read/mutate/fork helpers

Custom session stores

Implement SessionStore for external persistence. Stores can opt into session listing by overriding supports_list_sessions() and list_sessions(...).

use async_trait::async_trait;
use claude_agent_sdk::{
    Result, SessionKey, SessionListSubkeysKey, SessionStore, SessionStoreEntry,
    SessionStoreListEntry,
};

struct MyStore;

#[async_trait]
impl SessionStore for MyStore {
    async fn append(&self, key: SessionKey, entries: Vec<SessionStoreEntry>) -> Result<()> {
        println!("append {} entries to {key:?}", entries.len());
        Ok(())
    }

    async fn load(&self, key: SessionKey) -> Result<Option<Vec<SessionStoreEntry>>> {
        println!("load {key:?}");
        Ok(None)
    }

    async fn delete(&self, key: SessionKey) -> Result<()> {
        println!("delete {key:?}");
        Ok(())
    }

    async fn list_subkeys(&self, key: SessionListSubkeysKey) -> Result<Vec<String>> {
        println!("list subkeys for {key:?}");
        Ok(Vec::new())
    }

    fn supports_list_sessions(&self) -> bool {
        true
    }

    async fn list_sessions(&self, _project_key: &str) -> Result<Vec<SessionStoreListEntry>> {
        Ok(Vec::new())
    }
}

Error handling

Most functions return claude_agent_sdk::Result<T>, an alias for Result<T, ClaudeSDKError>.

Main error variants include:

• CLIConnectionError
• CLINotFoundError
• ProcessError
• CLIJSONDecodeError
• MessageParseError
• generic SDK/configuration errors

Testing

Run the standard suite:

cargo test

Run feature verification:

cargo check --features otel

Run the ignored real CLI smoke test only when authenticated and prepared for possible API usage:

cargo test --test e2e_cli -- --ignored

Crate layout

• client - interactive client
• query - one-shot query helpers
• types - public SDK data types
• options - builder API
• mcp - SDK MCP server/tool helpers
• session_store - custom store trait and in-memory store
• sessions - local and store-backed session helpers
• internal - CLI args, transport, protocol, control, parsing, mirroring

License

MIT