Open Agent SDK (Rust)

Build AI agents in Rust, over OpenAI chat completions or Anthropic messages

What you can build:

• Copy editors that analyze manuscripts and track writing patterns
• Git commit generators that write meaningful commit messages
• Market analyzers that research competitors and summarize findings
• Code reviewers, data analysts, research assistants, and more

Why this SDK?

• Two protocols - one API over OpenAI chat completions or Anthropic messages, chosen per endpoint
• Local or hosted - run on your own hardware at no API cost and with no data leaving the machine, or point it at a vendor
• Control - pick your model (Qwen, Llama, Mistral, Claude, etc.)

How fast? From zero to working agent in under 5 minutes. Rust-native performance (zero-cost abstractions, no GC), fearless concurrency, with 569 active tests.

---

Overview

Open Agent SDK (Rust) provides a clean, streaming API for working with local and cloud model servers over two wire protocols: OpenAI chat completions and Anthropic messages, selected per endpoint. 100% feature parity with the Python SDK—complete with transport-boundary-safe SSE streaming, tool call aggregation, hooks, and automatic tool execution—built on Tokio for high-performance async I/O.

Streaming is tolerant of real-world servers. SSE events are buffered across arbitrary HTTP transport chunk boundaries, and anything still held is flushed when the stream ends — including when a server closes the connection or sends data: [DONE] without ever setting finish_reason, which llama.cpp, vLLM, and several local gateways do. Content is never silently dropped.

Every stream reports why it ended. The stream from query() terminates with exactly one StreamEvent::Finish carrying a FinishReason, so a response cut off at the token cap (Length) is distinguishable from one the model chose to end (Stop) and from a server that never said (Unspecified) — three cases that look identical from the content alone. On Client, the same information is available from client.finish_reason().

Supported Providers

The protocol is a property of the endpoint, set with .protocol(..) and defaulting to ApiProtocol::OpenAiChat.

ApiProtocol::OpenAiChat — POST {base_url}/chat/completions, bearer auth

• LM Studio - http://localhost:1234/v1
• Ollama - http://localhost:11434/v1
• llama.cpp server - OpenAI-compatible mode
• vLLM - OpenAI-compatible API
• Text Generation WebUI - OpenAI extension
• Any OpenAI-compatible local endpoint
• Cloud vendors - OpenAI, OpenRouter, z.ai (https://api.z.ai/api/coding/paas/v4)
• Local gateways proxying cloud models - e.g., Ollama or custom gateways that route to cloud providers

Note on LM Studio: LM Studio is particularly well-tested with this SDK and provides reliable OpenAI-compatible API support. If you're looking for a user-friendly local model server with excellent compatibility, LM Studio is highly recommended.

ApiProtocol::Anthropic — POST {base_url}/messages, x-api-key + anthropic-version

• Anthropic
• Moonshot Kimi for Coding - https://api.kimi.com/coding/v1
• MiniMax - https://api.minimax.io/anthropic/v1

Extended thinking arrives on the existing reasoning channel (StreamEvent::Reasoning, opt in with .include_reasoning(true)), and tool calls as ordinary ContentBlock::ToolUse blocks.

Third-party Anthropic-compatible endpoints are not uniform, and the SDK invents nothing on their behalf: api.kimi.com/coding/v1 requires max_tokens and answers a bare invalid_request_error 400 without it, and accepts no temperature but 1. Set both explicitly when an endpoint asks for them.

Not Supported (Use Official SDKs)

• Cloud provider SDKs - Bedrock, Vertex, Azure, etc. (proxied via local gateway is fine)

Quick Start

Installation

[dependencies]
open-agent-sdk = "0.10.0"
tokio = { version = "1", features = ["full"] }
futures = "0.3"
serde_json = "1.0"

For development:

git clone https://github.com/slb350/open-agent-sdk-rust.git
cd open-agent-sdk-rust
cargo build

Upgrading from 0.9.x

v0.10.0 has one breaking change, and the compiler does not catch it: the types are unchanged, only how many events carry the same text.

Text and reasoning arrive fragment by fragment. A response that used to arrive as one ContentBlock::Text at the end of the stream now arrives as one block per delta, in order, while the stream is still open — which is what the SDK has always claimed streaming meant. Code that concatenates what it receives needs no change. Code that read the first text block as the whole answer now reads a prefix of it:

// Before: happened to work, because there was only ever one text block.
if let Some(StreamEvent::Block(ContentBlock::Text(text))) = stream.next().await.transpose()? {
    println!("{}", text.text);   // now prints the first few tokens only
}

// After: join the fragments, printing them as they land.
let mut answer = String::new();
while let Some(event) = stream.next().await {
    if let StreamEvent::Block(ContentBlock::Text(text)) = event? {
        print!("{}", text.text);
        answer.push_str(&text.text);
    }
}

If you already collect blocks and want the old shape back, coalesce_text_blocks is the join the SDK applies internally, now exported:

use open_agent::coalesce_text_blocks;

let whole = coalesce_text_blocks(&collected);   // adjacent text joined, tool calls untouched

Conversation history is unaffected — the fragments are joined before the assistant turn is written, so the next request replays exactly what 0.9.x replayed. Tool calls are unaffected: their arguments are only valid JSON once the last fragment lands, so they still emit whole at the end of the stream, in ascending index order.

Upgrading from 0.8.x

v0.9.0 has one breaking change, and the compiler catches it.

AgentOptions::temperature() returns Option<f32>, and unset now means unset. It used to default to 0.7 and was always sent. A growing number of models reject the parameter outright — Anthropic's range stops at 1.0, and Moonshot's k3 answers only temperature 1 is allowed for this model with a 400 — so a client-invented default turns a working request into a hard error. None omits the field and the server decides, exactly as max_tokens has behaved since 0.7.0. The builder is unchanged: .temperature(0.2) still sets one. Callers that relied on the old default now pass it explicitly.

Reaching an Anthropic messages endpoint is additive:

let options = AgentOptions::builder()
    .model("k3")
    .base_url("https://api.kimi.com/coding/v1")
    .api_key(&key)
    .protocol(ApiProtocol::Anthropic)
    .max_tokens(200_000)   // this endpoint requires it
    .build()?;

Upgrading from 0.7.x

v0.8.0 had one breaking change, and the compiler caught it.

query() yields StreamEvent instead of ContentBlock. The stream needed room for something that is not content: the reason generation stopped. Every stream now ends with exactly one StreamEvent::Finish.

// Before (0.7.x)
while let Some(block) = stream.next().await {
    match block? {
        ContentBlock::Text(text) => print!("{}", text.text),
        _ => {}
    }
}

// After (0.8.0) — smallest possible edit
while let Some(event) = stream.next().await {
    match event?.into_block() {
        Some(ContentBlock::Text(text)) => print!("{}", text.text),
        _ => {}
    }
}

If you parse structured output, the reason you upgraded is the Finish event — match it rather than discarding it:

use open_agent::{ContentBlock, FinishReason, StreamEvent};

let mut answer = String::new();
while let Some(event) = stream.next().await {
    match event? {
        StreamEvent::Block(ContentBlock::Text(text)) => answer.push_str(&text.text),
        // Truncated at the token cap: the JSON is missing because generation ran out of
        // budget, not because the model refused. Retry with a larger cap.
        StreamEvent::Finish(FinishReason::Length) => return Err("truncated".into()),
        // The model finished and still did not produce JSON. Retrying will not help.
        StreamEvent::Finish(FinishReason::Stop) => {}
        // The server never said. Neither conclusion is available.
        StreamEvent::Finish(FinishReason::Unspecified) => {}
        _ => {}
    }
}

ContentStream was renamed EventStream. ToolCallAggregator was renamed StreamAccumulator — if you imported it directly, update the import; the API is otherwise unchanged. ContentBlock is unchanged — no new variant and no wire-shape change, so exhaustive matches over it and its serde representation still work.

Client is unaffected. client.receive() still yields ContentBlock; the finish reason is recorded on the client instead:

client.send("Reply with JSON.").await?;
while let Some(block) = client.receive().await? { /* ... */ }

if client.finish_reason().is_some_and(FinishReason::is_truncated) {
    // Retry with a larger max_tokens.
}

New in 0.8.0, with no action required: reasoning-model side channels (reasoning_content on DeepSeek, reasoning on OpenRouter) are now explicitly parsed and routed away from assistant text rather than dropped as unknown fields, so deliberation prose can never be spliced into a response you parse as JSON. Opt into seeing it with .include_reasoning(true), which surfaces it as StreamEvent::Reasoning and client.reasoning().

Upgrading from 0.6.x

v0.7.0 has two breaking changes. Most projects need no edits at all; the compiler catches the first, and the second is a behaviour change with no compile error.

1. Error::Api carries the HTTP status. It changed from a tuple variant to a struct variant, so any pattern match must be updated:

// Before (0.6.x)
if let Error::Api(msg) = &err { eprintln!("{msg}"); }

// After (0.7.0)
if let Error::Api { message, status } = &err {
    eprintln!("{message} (status: {status:?})");
}

Constructing errors is unchanged — Error::api(msg) still works and yields status: None. Use the new Error::api_status(status, msg) when you have a status code, because is_retryable_error classifies on the status and treats a statusless API error as permanent.

2. max_tokens is no longer defaulted to 4096. Leaving .max_tokens() unset now omits the field from the request so the server applies its own limit. This is a silent behaviour change: if you relied on the implicit cap, set it explicitly.

let options = AgentOptions::builder()
    .model("qwen2.5-32b-instruct")
    .base_url("http://localhost:1234/v1")
    .max_tokens(4096)  // add this to keep the old behaviour
    .build()?;

Leaving it unset is recommended for long-context and reasoning models, which a 4096-token client-side cap truncates mid-response.

Also fixed in 0.7.0, with no action required: streamed content is no longer discarded when a server ends its stream without ever sending finish_reason (llama.cpp, vLLM, and several local gateways do this), and 429 is now correctly treated as retryable.

Simple Query (LM Studio)

use open_agent::{query, AgentOptions, ContentBlock, FinishReason, StreamEvent};
use futures::StreamExt;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let options = AgentOptions::builder()
        .system_prompt("You are a professional copy editor")
        .model("qwen2.5-32b-instruct")
        .base_url("http://localhost:1234/v1")
        .temperature(0.1)
        .build()?;

    let mut stream = query("Analyze this text...", &options).await?;

    while let Some(event) = stream.next().await {
        match event? {
            StreamEvent::Block(ContentBlock::Text(text)) => print!("{}", text.text),
            // Always emitted, exactly once, as the last event.
            StreamEvent::Finish(FinishReason::Length) => eprintln!("\n[truncated]"),
            _ => {}
        }
    }

    Ok(())
}

Multi-Turn Conversation (Ollama)

use open_agent::{Client, AgentOptions, ContentBlock};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let options = AgentOptions::builder()
        .system_prompt("You are a helpful assistant")
        .model("qwen3:8b")
        .base_url("http://localhost:11434/v1")
        .build()?;

    let mut client = Client::new(options)?;

    client.send("What's the capital of France?").await?;

    while let Some(block) = client.receive().await? {
        match block {
            ContentBlock::Text(text) => {
                print!("{}", text.text);
            }
            ContentBlock::ToolUse(tool_use) => {
                println!("Tool used: {}", tool_use.name());
                // Execute tool and add result
                // client.add_tool_result(tool_use.id(), result)?;
            }
            _ => {}
        }
    }

    Ok(())
}

Function Calling with Tools

Define tools using the builder pattern for clean, type-safe function calling:

use open_agent::{tool, Client, AgentOptions, ContentBlock};
use serde_json::json;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Define tools
    let add_tool = tool("add", "Add two numbers")
        .param("a", "number")
        .param("b", "number")
        .build(|args| async move {
            let a = args["a"].as_f64().unwrap_or(0.0);
            let b = args["b"].as_f64().unwrap_or(0.0);
            Ok(json!({"result": a + b}))
        });

    // Enable automatic tool execution (recommended)
    let options = AgentOptions::builder()
        .system_prompt("You are a helpful assistant with access to tools.")
        .model("qwen2.5-32b-instruct")
        .base_url("http://localhost:1234/v1")
        .tool(add_tool)
        .auto_execute_tools(true)      // Tools execute automatically
        .max_tool_iterations(10)       // Safety limit for tool loops
        .build()?;

    let mut client = Client::new(options)?;
    client.send("What's 25 + 17?").await?;

    // Simply iterate - tools execute automatically!
    while let Some(block) = client.receive().await? {
        match block {
            ContentBlock::Text(text) => {
                println!("Response: {}", text.text);
            }
            _ => {}
        }
    }

    Ok(())
}

Advanced: Manual Tool Execution

For custom execution logic or result interception:

// Disable auto-execution
let options = AgentOptions::builder()
    .system_prompt("You are a helpful assistant with access to tools.")
    .model("qwen2.5-32b-instruct")
    .base_url("http://localhost:1234/v1")
    .tool(add_tool.clone())
    .auto_execute_tools(false)  // Manual mode
    .build()?;

let mut client = Client::new(options)?;
client.send("What's 25 + 17?").await?;

while let Some(block) = client.receive().await? {
    match block {
        ContentBlock::ToolUse(tool_use) => {
            // You execute the tool manually
            let result = add_tool.execute(tool_use.input().clone()).await?;

            // Return result to agent
            client.add_tool_result(tool_use.id(), result)?;

            // Continue conversation
            client.send("").await?;
        }
        ContentBlock::Text(text) => {
            print!("{}", text.text);
        }
        _ => {}
    }
}

Key Features:

• Automatic execution - Tools run automatically with safety limits
• Type-safe schemas - Automatic JSON schema generation from parameters
• Both protocols - The same tool definitions serve OpenAI function calling and Anthropic tool use
• Clean builder API - Fluent API for tool definition
• Hook integration - PreToolUse/PostToolUse hooks work in both modes

See examples/calculator_tools.rs and examples/auto_execution_demo.rs for complete examples.

Multimodal Vision Support

Send images alongside text to vision-capable models like llava, qwen-vl, or minicpm-v. The SDK formats images for whichever protocol the endpoint speaks.

Simple Image + Text

use open_agent::{Client, Message, MessageRole, ContentBlock, TextBlock, ImageBlock, ImageDetail};

// From URL
let msg = Message::user_with_image(
    "What's in this image?",
    "https://example.com/photo.jpg"
)?;
client.send_message(msg).await?;

// From local file path (NEW!)
let msg = Message::new(
    MessageRole::User,
    vec![
        ContentBlock::Text(TextBlock::new("Describe this photo")),
        ContentBlock::Image(ImageBlock::from_file_path("/path/to/photo.jpg")?),
    ],
);
client.send_message(msg).await?;

// From base64 data
let msg = Message::user_with_base64_image(
    "Describe this diagram",
    base64_data,
    "image/png"
)?;
client.send_message(msg).await?;

// Control detail level for token costs
let msg = Message::user_with_image_detail(
    "Analyze the fine details",
    "https://example.com/diagram.png",
    ImageDetail::High  // Low: ~85 tokens, High: variable, Auto: default
)?;
client.send_message(msg).await?;

Supported Image Sources:

• ImageBlock::from_url(url) - HTTPS/HTTP URLs or data URIs (e.g., data:image/png;base64,...)
• ImageBlock::from_file_path(path) - Local filesystem (automatically encodes as base64)
  • Supports: .jpg, .jpeg, .png, .gif, .webp, .bmp, .svg
  • MIME type inferred from file extension
  • File is read and encoded automatically
• ImageBlock::from_base64(data, mime) - Manual base64 with explicit MIME type

Token Cost Management

Control image processing costs using ImageDetail levels:

• ImageDetail::Low - Lower resolution (typically more cost-effective)
• ImageDetail::High - Higher resolution (typically more detailed analysis)
• ImageDetail::Auto - Model decides (balanced default)

⚠️ Token Costs Vary by Model:

OpenAI's Vision API uses ~85 tokens (Low) and variable tokens based on dimensions (High), but local models may have completely different token costs—or no token costs for images at all. The ImageDetail setting may even be ignored by some models.

Always benchmark your specific model instead of relying on OpenAI's published values for capacity planning.

Complex Multi-Image Messages

use open_agent::{Message, MessageRole, ContentBlock, TextBlock, ImageBlock, ImageDetail};

let msg = Message::new(
    MessageRole::User,
    vec![
        ContentBlock::Text(TextBlock::new("Compare these images:")),
        ContentBlock::Image(
            ImageBlock::from_url("https://example.com/before.jpg")?
                .with_detail(ImageDetail::Low)
        ),
        ContentBlock::Image(
            ImageBlock::from_url("https://example.com/after.jpg")?
                .with_detail(ImageDetail::Low)
        ),
    ],
);

Key Features:

• send_message() API - Send pre-built messages with images via client.send_message(msg).await?
• Automatic serialization - Images converted to OpenAI Vision or Anthropic image blocks (ImageDetail has no Anthropic equivalent and is dropped there)
• Multiple sources - URLs, local file paths, or base64 data
• Backward compatible - Text-only messages still work with send("text")
• Data URIs supported - Base64-encoded images transmitted seamlessly
• Token cost control - Choose detail level based on use case

See examples/vision_example.rs for comprehensive working examples including local file paths.

Context Management

Local models have fixed context windows (typically 8k-32k tokens). The SDK provides utilities for manual history management—no silent mutations, you stay in control.

Token Estimation & Truncation

use open_agent::{Client, AgentOptions, estimate_tokens, is_approaching_limit, truncate_messages};

let mut client = Client::new(options)?;

// Long conversation...
for i in 0..50 {
    client.send(&format!("Question {}", i)).await?;
    while let Some(block) = client.receive().await? {
        // Process blocks
        let _ = block;
    }
}

// Check token usage
let tokens = estimate_tokens(client.history());
println!("Context size: ~{} tokens", tokens);

// Check if approaching limit (margin = 0.8 means warn at 80% of limit)
if is_approaching_limit(client.history(), 32000, 0.8) {
    println!("Warning: approaching context limit");
}

// Manually truncate when needed
if tokens > 28000 {
    let truncated = truncate_messages(client.history(), 10, true);
    *client.history_mut() = truncated;
}

Recommended Patterns

1. Stateless Agents (Best for single-task agents):

// Process each task independently - no history accumulation
for task in tasks {
    let mut client = Client::new(options.clone());
    client.send(&task).await?;
    // Client dropped, fresh context for next task
}

2. Manual Truncation (At natural breakpoints):

use open_agent::truncate_messages;

let mut client = Client::new(options)?;
for task in tasks {
    client.send(&task).await?;
    // Truncate after each major task
    let truncated = truncate_messages(client.history(), 5, false);
    *client.history_mut() = truncated;
}

3. External Memory (RAG-lite for research agents):

// Store important facts in database, keep conversation context small
let mut database = HashMap::new();
let mut client = Client::new(options)?;

client.send("Research topic X").await?;
// Save response to database
database.insert("topic_x", extract_facts(&response));

// Clear history, query database when needed
let truncated = truncate_messages(client.history(), 0, false);
*client.history_mut() = truncated;

Why Manual?

The SDK intentionally does not auto-compact history because:

• Domain-specific needs: Copy editors need different strategies than research agents
• Token accuracy varies: Each model family has different tokenizers
• Risk of breaking context: Silently removing messages could break tool chains
• Natural limits exist: Compaction doesn't bypass model context windows

See examples/context_management.rs for complete patterns and usage.

Lifecycle Hooks

Monitor and control agent behavior at key execution points with zero-cost Rust hooks.

Quick Example

use open_agent::{
    AgentOptions, Client, Hooks,
    PreToolUseEvent, PostToolUseEvent,
    HookDecision,
};

// Security gate - block dangerous operations
let hooks = Hooks::new()
    .add_pre_tool_use(|event| async move {
        if event.tool_name == "delete_file" {
            return Some(HookDecision::block("Delete operations require approval"));
        }
        Some(HookDecision::continue_())
    })
    .add_post_tool_use(|event| async move {
        // Audit logger - track all tool executions
        println!("Tool executed: {} -> {:?}", event.tool_name, event.tool_result);
        None
    });

// Register hooks in AgentOptions
let options = AgentOptions::builder()
    .system_prompt("You are a helpful assistant")
    .model("qwen2.5-32b-instruct")
    .base_url("http://localhost:1234/v1")
    .hooks(hooks)
    .build()?;

let mut client = Client::new(options)?;

Hook Types

PreToolUse - Fires before tool execution

• Block operations: Return Some(HookDecision::block(reason))
• Modify inputs: Return Some(HookDecision::modify_input(json!({}), reason))
• Allow: Return Some(HookDecision::continue_())

PostToolUse - Fires after the tool completes and before the final result is committed

• Observational (tool already executed)
• Use for audit logging, metrics, result validation
• Return None or Some(HookDecision::...)

Every hook event exposes history as Vec<serde_json::Value>, with one structured JSON object per internal Message (role plus typed content blocks). Prompt and pre-tool snapshots contain history up to that lifecycle point; post-tool snapshots also include the completed tool call and its unmodified result.

UserPromptSubmit - Fires before sending prompt to API

• Block prompts: Return Some(HookDecision::block(reason))
• Modify prompts: Return Some(HookDecision::modify_prompt(text, reason))
• Allow: Return Some(HookDecision::continue_())

Common Patterns

Pattern 1: Redirect to Sandbox

hooks.add_pre_tool_use(|event| async move {
    if event.tool_name == "file_writer" {
        let path = event.tool_input.get("path")
            .and_then(|v| v.as_str())
            .unwrap_or("");

        if !path.starts_with("/tmp/") {
            let safe_path = format!("/tmp/sandbox/{}", path.trim_start_matches('/'));
            let mut modified = event.tool_input.clone();
            modified["path"] = json!(safe_path);
            return Some(HookDecision::modify_input(modified, "Redirected to sandbox"));
        }
    }
    Some(HookDecision::continue_())
})

Pattern 2: Compliance Audit Log

let audit_log = Arc::new(Mutex::new(Vec::new()));
let log_clone = audit_log.clone();

// Note: add_post_tool_use consumes and returns Hooks (builder pattern) — always rebind
let hooks = hooks.add_post_tool_use(move |event| {
    let log = log_clone.clone();
    async move {
        log.lock().unwrap().push(format!(
            "{} -> {:?}",
            event.tool_name,
            event.tool_result
        ));
        None
    }
});

Hook Execution Flow

• Hooks run sequentially in the order registered
• First non-None decision wins (short-circuit behavior)
• Hooks run inline on async runtime (spawn tasks for heavy work)
• Works with both Client and query() function

See examples/hooks_example.rs and examples/multi_tool_agent.rs for comprehensive patterns.

Interrupt Capability

Cancel long-running operations cleanly without corrupting client state. Perfect for timeouts, user cancellations, or conditional interruptions.

Interrupt Quick Example

use open_agent::{Client, AgentOptions};
use tokio::time::{timeout, Duration};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let options = AgentOptions::builder()
        .system_prompt("You are a helpful assistant.")
        .model("qwen2.5-32b-instruct")
        .base_url("http://localhost:1234/v1")
        .build()?;

    let mut client = Client::new(options)?;
    client.send("Write a detailed 1000-word essay...").await?;

    // Timeout after 5 seconds
    match timeout(Duration::from_secs(5), async {
        while let Some(block) = client.receive().await? {
            // Process blocks
            let _ = block;
        }
        Ok::<_, Box<dyn std::error::Error>>(())
    }).await {
        Ok(_) => println!("Completed"),
        Err(_) => {
            client.interrupt();  // Clean cancellation
            println!("Operation timed out!");
        }
    }

    // Client is still usable after interrupt
    client.send("Short question?").await?;
    // Continue using client...

    Ok(())
}

Common Interrupt Patterns

1. Conditional Interruption

let mut full_text = String::new();
while let Some(block) = client.receive().await? {
    if let ContentBlock::Text(text) = block {
        full_text.push_str(&text.text);
        if full_text.contains("error") {
            client.interrupt();
            break;
        }
    }
}

2. Concurrent Cancellation

use std::sync::atomic::Ordering;

let interrupt_handle = client.interrupt_handle();
let cancel_task = tokio::spawn(async move {
    tokio::time::sleep(Duration::from_secs(2)).await;
    interrupt_handle.store(true, Ordering::SeqCst);
});

while let Some(block) = client.receive().await? {
    // Process blocks until cancellation is observed
    let _ = block;
}

cancel_task.await?;

How It Works

When you call client.interrupt():

1. Atomic signal - A thread-safe flag tells the receive loop to stop
2. Stream cleanup - receive() observes the flag, drops the active stream, and returns Ok(None)
3. Clean history - Partial manual responses are discarded instead of committing incomplete assistant messages
4. Idempotent - Safe to call multiple times
5. Cross-task safe - interrupt_handle() lets another task cancel without locking the Client

See examples/interrupt_demo.rs for comprehensive patterns.

Practical Examples

Example agents demonstrating real-world usage:

Git Commit Agent

examples/git_commit_agent.rs

Analyzes your staged git changes and writes professional commit messages following conventional commit format.

# Stage your changes
git add .

# Run the agent
cargo run --example git_commit_agent

# Output:
# Found staged changes in 3 file(s)
# Analyzing changes and generating commit message...
#
# Suggested commit message:
# feat(auth): Add OAuth2 integration with refresh tokens
#
# - Implement token refresh mechanism
# - Add secure cookie storage for tokens
# - Update login flow to support OAuth2 providers

Features:

• Analyzes diff to determine commit type (feat/fix/docs/etc)
• Writes clear, descriptive commit messages
• Follows conventional commit standards

Log Analyzer Agent

examples/log_analyzer_agent.rs

Intelligently analyzes application logs to identify patterns, errors, and provide actionable insights.

# Analyze a log file
cargo run --example log_analyzer_agent -- /var/log/app.log

Features:

• Automatic error pattern detection
• Time-based analysis (peak error times)
• Root cause suggestions
• Supports multiple log formats

Why These Examples?

These agents demonstrate:

• Practical Value: Solve real problems developers face daily
• Tool Integration: Show how to integrate with system commands (git, file I/O)
• Structured Output: Parse and format LLM responses for actionable results
• Privacy-First: Keep your code and logs local while getting AI assistance

Why Not Just Use OpenAI Client?

Without open-agent-sdk (raw reqwest):

use reqwest::Client;

let client = Client::new();
let response = client
    .post("http://localhost:1234/v1/chat/completions")
    .json(&json!({
        "model": "qwen2.5-32b-instruct",
        "messages": [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt}
        ],
        "stream": true
    }))
    .send()
    .await?;

// Complex parsing of SSE chunks
// Extract delta content
// Handle tool calls manually
// Track conversation state yourself

With open-agent-sdk:

use open_agent::{query, AgentOptions};

let options = AgentOptions::builder()
    .system_prompt(system_prompt)
    .model("qwen2.5-32b-instruct")
    .base_url("http://localhost:1234/v1")
    .build()?;

let mut stream = query(user_prompt, &options).await?;
// Clean message types (TextBlock, ToolUseBlock)
// Automatic streaming and tool call handling
// Terminating StreamEvent::Finish tells you why generation stopped

Value: Familiar patterns + Less boilerplate + Rust performance

Why Rust?

Performance: Zero-cost abstractions mean no runtime overhead. Streaming responses with Tokio delivers throughput comparable to C/C++ while maintaining memory safety.

Safety: Compile-time guarantees prevent data races, null pointer dereferences, and buffer overflows. Your agents won't crash from memory issues.

Concurrency: Fearless concurrency with async/await lets you run multiple agents or handle hundreds of concurrent requests without fear of race conditions.

Production Ready: Strong type system catches bugs at compile time. Comprehensive error handling with Result types. No surprises in production.

Small Binaries: Standalone executables under 10MB. Deploy anywhere without runtime dependencies.

API Reference

AgentOptions

AgentOptions::builder()
    .system_prompt(str)                  // System prompt
    .model(str)                          // Model name (required)
    .base_url(str)                       // Endpoint URL; the path comes from .protocol() (required)
    .tool(Tool)                          // Add a single tool for function calling
    .tools(Vec<Tool>)                    // Add multiple tools at once
    .hooks(Hooks)                        // Lifecycle hooks for monitoring/control
    .auto_execute_tools(bool)            // Enable automatic tool execution
    .max_tool_iterations(u32)            // Max tool calls per query in auto mode
    .max_tokens(u32)                     // Tokens to generate (unset: omitted, server decides); getter returns Option<u32>
    .max_turns(u32)                      // Max conversation turns (default: 1)
    .temperature(f32)                    // Sampling temperature (unset: omitted, server decides)
    .protocol(ApiProtocol)               // Wire protocol (default: ApiProtocol::OpenAiChat)
    .timeout(u64)                        // Request timeout in seconds (default: 60)
    .api_key(str)                        // API key (default: "not-needed")
    .include_reasoning(bool)             // Surface reasoning as StreamEvent::Reasoning (default: false)
    .build()?

query()

Simple single-turn query function.

pub async fn query(prompt: &str, options: &AgentOptions) -> Result<EventStream>

// where
pub type EventStream = Pin<Box<dyn Stream<Item = Result<StreamEvent>> + Send>>;

Returns a stream yielding StreamEvent items. Use futures::StreamExt to iterate.

StreamEvent and FinishReason

pub enum StreamEvent {
    Block(ContentBlock),   // Assistant text, or a fully assembled tool call
    Reasoning(String),     // Chain of thought; only with .include_reasoning(true)
    Finish(FinishReason),  // Exactly once per stream, always last
}

pub enum FinishReason {
    Stop,               // "stop"           — completed naturally
    Length,             // "length"         — cut off at the token limit
    ToolCalls,          // "tool_calls"     — finished in order to call tools
    ContentFilter,      // "content_filter" — halted by a content filter
    Other(String),      // anything else, preserved verbatim
    MaxToolIterations,  // the SDK's auto-execution loop hit max_tool_iterations
    Unspecified,        // the stream ended and the server never said
}

Both are #[non_exhaustive]; match with a _ arm. StreamEvent provides as_block(), into_block(), as_text(), as_reasoning(), and finish_reason(). FinishReason provides from_wire(), as_str(), is_truncated(), and Display.

Unspecified is not an error — it is the normal behaviour of llama.cpp, vLLM, and several local gateways, which stream content and then close without setting finish_reason. It is kept distinct from Stop because "the model finished" and "the SDK has no information" call for different handling.

MaxToolIterations is the one variant that does not come from a server: in auto-execution mode the SDK, not the model, ends the run when it hits max_tool_iterations. It is reported only by client.finish_reason() and never appears in a StreamEvent::Finish.

Client

Multi-turn conversation client with tool monitoring.

let mut client = Client::new(options)?;
client.send(prompt).await?;

while let Some(block) = client.receive().await? {
    // Process ContentBlock items
    let _ = block;
}

Additional Client methods:

// Send a pre-built Message (e.g., with images)
client.send_message(msg).await?;

// Access the AgentOptions this client was created with
let opts = client.options();

// Clear conversation history (resets to system prompt only)
client.clear_history();

// Look up a registered tool by name
if let Some(t) = client.get_tool("my_tool") { /* ... */ }

// Obtain a shareable interrupt handle (Arc<AtomicBool>) for use across tasks
let handle = client.interrupt_handle();

// Why the most recent stream stopped; None until one completes, reset on the next send()
if let Some(reason) = client.finish_reason() { println!("stopped: {reason}"); }

// Reasoning captured from the most recent turn (requires .include_reasoning(true));
// accumulates across every round of an auto-execution tool loop
if let Some(reasoning) = client.reasoning() { println!("thought: {reasoning}"); }

MessageRole

Who sent a message. Used when constructing Message values directly.

use open_agent::MessageRole;

MessageRole::System     // Establishes context and instructions
MessageRole::User       // Input from the human or calling application
MessageRole::Assistant  // Response from the AI model
MessageRole::Tool       // Results from tool/function execution

Message

Pre-built message values (for client.send_message()). Convenience constructors:

use open_agent::{Message, MessageRole, ContentBlock, TextBlock};

// Build a message manually (any role)
Message::new(role: MessageRole, content: Vec<ContentBlock>) -> Self

// Convenience constructors — all return Self (infallible):
Message::user(text: &str) -> Self
Message::assistant(text: &str) -> Self
Message::system(text: &str) -> Self
Message::user_with_blocks(blocks: Vec<ContentBlock>) -> Self

// Vision constructors — return Result<Self>:
Message::user_with_image(text: &str, image_url: &str) -> Result<Self>
Message::user_with_image_detail(text: &str, image_url: &str, detail: ImageDetail) -> Result<Self>
Message::user_with_base64_image(text: &str, base64_data: &str, mime: &str) -> Result<Self>

Message Types

• ContentBlock::Text(TextBlock) - Text content from model
• ContentBlock::Image(ImageBlock) - Image content (for vision models)
• ContentBlock::ToolUse(ToolUseBlock) - Tool calls from model
• ContentBlock::ToolResult(ToolResultBlock) - Tool execution results

Tool System

use open_agent::tool;

let my_tool = tool("name", "description")
    .param("param_name", "type")
    .build(|args| async move {
        // Tool implementation
        Ok(json!({"result": value}))
    });

For full JSON Schema control, use .schema() instead of chaining .param() calls:

let my_tool = tool("name", "description")
    .schema(json!({
        "type": "object",
        "properties": { "x": { "type": "number" } },
        "required": ["x"]
    }))
    .build(|args| async move { Ok(json!({})) });

ToolBuilder

The tool() function returns a ToolBuilder for fluent construction of tool definitions:

use open_agent::{tool, ToolBuilder, Tool};

let t: Tool = tool("name", "description")
    .param("arg", "string")
    .build(|args| async move { Ok(json!({})) });

Provider Configuration

Helper types and functions for mapping provider names to their default endpoints:

use open_agent::{Provider, get_base_url, get_model};

// get_base_url(provider: Option<Provider>, fallback: Option<&str>) -> String
let url = get_base_url(Some(Provider::LMStudio), None);   // http://localhost:1234/v1
let url_with_fallback = get_base_url(None, Some("http://localhost:8080/v1"));

// get_model(fallback: Option<&str>, prefer_env: bool) -> Option<String>
let model = get_model(Some("qwen2.5-32b"), false);  // use provided model
let env_model = get_model(None, true);              // prefer OPEN_AGENT_MODEL env var

Wire Types

Low-level serialization types matching each protocol's request and streaming format, exported for callers that need to name what goes over the wire:

// OpenAI chat completions
use open_agent::{
    OpenAIContent, OpenAIContentPart, OpenAIFunction, OpenAIMessage, OpenAIRequest, OpenAIToolCall,
};

// Anthropic messages
use open_agent::{
    AnthropicBlockStart, AnthropicDelta, AnthropicErrorBody, AnthropicEvent, AnthropicMessage,
    AnthropicMessageDelta, AnthropicRequest, anthropic_finish_reason,
};

AnthropicRequest::from_openai takes an OpenAIRequest, which is why the OpenAI request half is exported alongside the Anthropic types. anthropic_finish_reason maps Anthropic stop reasons onto FinishReason, and query()/Client apply it for you.

Error and Result Types

use open_agent::{Error, Result};

Error is the SDK's unified error type; Result<T> is an alias for std::result::Result<T, Error>.

Variant	Meaning
Http(reqwest::Error)	Transport failure — connection refused, DNS, TLS, network timeout
Json(serde_json::Error)	Serialization or deserialization failure
Config(String)	Invalid configuration caught by AgentOptions::build()
Api { status: Option<u16>, message: String }	Error response from the model server
Stream(String)	SSE parsing or stream processing failure
Tool(String)	Tool execution or registration failure
InvalidInput(String)	User-provided input failed validation
Timeout	Request exceeded the configured timeout
Other(String)	Anything else

Api carries the HTTP status as structured data so retry logic never has to parse the message text:

use open_agent::Error;

// From an HTTP error response — this is what the client constructs internally
let err = Error::api_status(429, "Rate limit exceeded");
assert_eq!(err.status_code(), Some(429));
assert_eq!(err.to_string(), "API error 429: Rate limit exceeded");

// Without a status
let err = Error::api("Model 'gpt-4' not found");
assert_eq!(err.status_code(), None);

status_code() returns None for every non-Api variant, so it is safe to call on any error.

Newtype Wrappers

Strong-typed wrappers used internally by AgentOptions and exported for external use:

use open_agent::{BaseUrl, ModelName, Temperature};

Retry Module

Exponential-backoff retry utilities, exported as a public module:

use open_agent::retry::{RetryConfig, retry_with_backoff, retry_with_backoff_conditional, is_retryable_error};

// Configure retry behavior (builder pattern)
let config = RetryConfig::default()          // 3 attempts, exponential backoff
    .max_attempts(5)
    .initial_delay_ms(100)
    .max_delay_ms(5000)
    .backoff_multiplier(2.0);

// Retry any async operation
let result = retry_with_backoff(config.clone(), || async {
    some_fallible_operation().await
}).await?;

// Retry only transient failures; anything else fails on the first attempt
let result = retry_with_backoff_conditional(config, || async {
    some_fallible_operation().await
}).await?;

// Check if an SDK error is worth retrying
let retryable = is_retryable_error(&some_error);

is_retryable_error treats network errors, timeouts, and stream errors as transient. API errors are classified on Error::status_code(), which reads the status Error::Api carries as structured data; the retryable set is 408, 429, 500, 502, 503, 504, 529. Everything else — including API errors raised without a status — is non-retryable, so a 400 Bad Request fails immediately rather than burning the full attempt budget.

A mid-stream Anthropic error event arrives on a response that already returned 200, so it carries no status of its own. The SDK maps the two transient kinds onto the statuses they would have had earlier — overloaded_error to 529 and rate_limit_error to 429, with api_error to 500 — which is what lets retry_with_backoff see them as retryable.

use open_agent::Error;

let err = Error::api_status(429, "Rate limit exceeded");   // status: Some(429)
assert_eq!(err.status_code(), Some(429));

let err = Error::api("Model 'gpt-4' not found");            // status: None
assert_eq!(err.status_code(), None);

Prelude Import

For convenience, import the most commonly used types at once:

use open_agent::prelude::*;

Hook Name Constants

String constants for hook event types are exported for use in custom registries:

use open_agent::{HOOK_PRE_TOOL_USE, HOOK_POST_TOOL_USE, HOOK_USER_PROMPT_SUBMIT};

Context Utilities

use open_agent::{estimate_tokens, is_approaching_limit, truncate_messages};

// Estimate tokens in message history (character-based approximation)
let tokens = estimate_tokens(client.history());

// Check if approaching a context limit (margin=0.8 means 80% of limit)
let near_limit = is_approaching_limit(client.history(), 32000, 0.8);

// Truncate history, keeping the last N messages (preserve_system=true keeps system prompt)
let truncated = truncate_messages(client.history(), 10, true);

Recommended Models

Local models (LM Studio, Ollama, llama.cpp):

• GPT-OSS-120B - Best in class for speed and quality
• Qwen 3 30B - Excellent instruction following, good for most tasks
• GPT-OSS-20B - Solid all-around performance
• Mistral 7B - Fast and efficient for simple agents

Cloud-proxied via local gateway:

• kimi-k2:1t-cloud - Tested and working via Ollama gateway
• deepseek-v3.1:671b-cloud - High-quality reasoning model
• qwen3-coder:480b-cloud - Code-focused models

Project Structure

open-agent-sdk-rust/
├── src/
│   ├── client.rs          # Public client module docs/imports and fragment orchestration
│   ├── client/            # Query, send, send_message, setup, streaming, receive, history, state, and tests
│   ├── config.rs          # Provider helpers (Provider, get_base_url, get_model)
│   ├── context.rs         # Token estimation and truncation
│   ├── error.rs           # Error types
│   ├── hooks.rs           # Public lifecycle-hook module orchestration
│   ├── hooks/             # Hook events, decisions, handlers, registry, and tests
│   ├── lib.rs             # Public exports and prelude module
│   ├── retry.rs           # Retry logic with exponential backoff
│   ├── retry/             # Retry unit tests
│   ├── tools.rs           # Public tool module orchestration
│   ├── tools/             # Tool, schema, builder, handler, factory, and tests
│   ├── types.rs           # Public core-type module orchestration
│   ├── types/             # Options, messages, images, OpenAI + Anthropic wire types, ApiProtocol
│   ├── utils.rs           # SSE parsing, stream accumulation, and the shared stream driver
│   └── utils/             # accumulator.rs + anthropic_accumulator.rs (wire decoding),
│                          # buffers.rs (the shared drain), coalesce.rs (text joining for
│                          # history), driver.rs, sse.rs
├── examples/
│   ├── simple_query.rs              # Basic streaming query
│   ├── anthropic_query.rs           # Anthropic messages endpoint via ApiProtocol
│   ├── calculator_tools.rs          # Function calling (manual mode)
│   ├── auto_execution_demo.rs       # Automatic tool execution
│   ├── multi_tool_agent.rs          # Production agent with 5 tools and hooks
│   ├── hooks_example.rs             # Lifecycle hooks patterns
│   ├── context_management.rs        # Context management patterns
│   ├── interrupt_demo.rs            # Interrupt capability patterns
│   ├── git_commit_agent.rs          # Production: Git commit generator
│   ├── log_analyzer_agent.rs        # Production: Log analyzer
│   ├── advanced_patterns.rs         # Retry logic and concurrent requests
│   ├── vision_example.rs            # Multimodal: URLs, local files, base64
│   ├── vision_api_demo.rs           # Vision API walkthrough
│   └── test_tool_serialization.rs   # Tool call serialization verification
├── benches/
│   └── performance.rs               # Criterion benchmarks (token estimation, history ops)
├── tests/
│   ├── integration_tests.rs         # Core integration tests
│   ├── regression_incremental_streaming_test.rs  # Per-fragment delivery; joined history
│   ├── advanced_integration_test.rs
│   ├── anthropic_protocol_test.rs   # Path, headers, body, and event vocabulary per protocol
│   ├── auto_execution_test.rs
│   ├── backward_compatibility_test.rs
│   ├── ci_workflow_policy_test.rs   # GitHub CI runner, coverage, mutation, and security guards
│   ├── client_image_serialization_test.rs
│   ├── config_env_test.rs           # get_model env resolution (own process: mutates env)
│   ├── context_estimation_test.rs   # Token arithmetic and truncation boundaries
│   ├── debug_logging_test.rs
│   ├── defensive_validation_test.rs
│   ├── edge_cases_test.rs
│   ├── hooks_history_snapshot_test.rs
│   ├── hooks_integration_test.rs
│   ├── image_serialization_test.rs
│   ├── package_manifest_test.rs     # Package exclusion coverage
│   ├── regression_finish_reason_test.rs
│   ├── regression_max_tokens_test.rs
│   ├── regression_reasoning_channel_test.rs
│   ├── regression_retry_classification_test.rs
│   ├── regression_stream_flush_test.rs
│   ├── security_bypass_test.rs
│   ├── send_message_test.rs         # Manual-mode history regression (v0.6.2)
│   ├── source_file_size_test.rs     # Repository Rust hard-limit guard
│   ├── tool_call_content_test.rs    # Tool call serialization tests
│   └── common/mod.rs                # Shared wiremock SSE harness and block helpers
├── scripts/
│   ├── mutants-common.sh            # The one definition of the results directory
│   ├── mutants-run.sh               # Owns the verdict (missed.txt); called by the hook and CI
│   ├── mutants-remote.sh            # rsync + ssh to a build host, falls back loudly
│   └── mutants-staged.sh            # Staged-diff scope, through mutants-remote.sh
├── .githooks/
│   └── pre-commit                   # fmt, clippy, tests, and a --in-diff cargo-mutants sweep
├── .github/
│   ├── dependabot.yml               # Grouped weekly Cargo dependency updates
│   └── workflows/
│       ├── ci.yml                   # GitHub CI (fmt, clippy, MSRV, Linux/macOS stable + beta matrix, security audit, mutation sweep, docs, LLVM Tarpaulin coverage, benchmarks)
│       └── scheduled-audit.yml      # Scheduled dependency audit
├── .markdownlint.json               # Markdown lint rules (disable MD013, allow duplicate sibling headings)
├── Cargo.toml
├── Cargo.lock
├── CHANGELOG.md
└── README.md

Examples

Production Agents

• git_commit_agent.rs – Analyzes git diffs and writes professional commit messages
• log_analyzer_agent.rs – Parses logs, finds patterns, suggests fixes
• multi_tool_agent.rs – Complete production setup with 5 tools, hooks, and auto-execution

Core SDK Usage

• simple_query.rs – Minimal streaming query (simplest quickstart)
• anthropic_query.rs – Same query against an Anthropic messages endpoint via ApiProtocol
• calculator_tools.rs – Manual tool execution pattern
• auto_execution_demo.rs – Automatic tool execution pattern
• vision_example.rs – Multimodal image support (URLs, local files, base64)
• vision_api_demo.rs – Vision API walkthrough with token cost notes
• hooks_example.rs – Lifecycle hooks patterns (security gates, audit logging)
• context_management.rs – Manual history management patterns
• interrupt_demo.rs – Interrupt capability patterns (timeout, conditional, concurrent)
• advanced_patterns.rs – Retry logic and concurrent request handling
• test_tool_serialization.rs – Verifies tool call serialization (see examples/test_tool_serialization.rs)

Documentation

• API Documentation
• Python SDK - Reference implementation
• Examples - Comprehensive usage examples

Testing

# Run all tests
cargo test

# Run with output
cargo test -- --nocapture

# Run specific test
cargo test test_agent_options_builder

# Mutation sweep (must report zero survivors)
cargo mutants --no-shuffle -j 4

Test Coverage:

• 248 unit tests (lib)
• 157 active integration tests across 26 test files
• 164 active doctests

Total: 569 active unit, integration, and documentation tests

Mutation testing is part of the gate, not an optional extra: a green suite proves the tests ran, not that they would notice if the code were wrong. CI runs the full sweep on every push. To run the same check before each commit:

git config core.hooksPath .githooks

The hook runs cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test, and a cargo mutants --in-diff sweep scoped to the staged Rust changes. Both the hook and CI reach their verdict through scripts/mutants-run.sh, which reads missed.txt rather than the exit code — cargo-mutants reports a timeout in preference to a survivor, so a run with one of each would otherwise look like a timeout. The current sweep is 231 mutants, 0 missed.

Requirements

• Rust 1.85+
• Tokio 1.50+ (async runtime)
• serde, serde_json (serialization)
• reqwest (HTTP client)
• futures, tokio-stream (async streams)
• eventsource-stream (SSE parsing)
• async-trait (async trait support)
• thiserror 2.0 + anyhow 1.0.103+ (error handling)
• log 0.4.29+ (logging)
• base64 0.23 (multimodal image encoding)
• wiremock 0.6 (dev-only: HTTP mocking for streaming and wire-format tests)
• cargo-mutants 27.1.0 (dev-only: mutation testing gate)
• rand (retry jitter)

License

MIT License - see LICENSE for details.

Acknowledgments

• Rust port of open-agent-sdk Python library
• API design inspired by claude-agent-sdk
• Built for local/open-source LLM enthusiasts

Repository Hosting

GitHub is the canonical repository and CI/release host. Any family Gitea copy is a passive Git mirror and does not run a separate required Actions pipeline.

---

Status: v0.10.0 - Incremental text and reasoning delivery, plus from v0.9.0 the Anthropic messages protocol alongside OpenAI chat completions selected per endpoint with ApiProtocol, extended thinking on the existing reasoning channel, omittable temperature, plus from v0.8.0 finish reasons surfaced on every stream, explicit reasoning-channel separation, end-of-stream flushing for servers that omit finish_reason, structured Error::Api with status-based retry classification, no implicit max_tokens cap, a mandatory mutation-testing gate, plus transport-boundary-safe SSE streaming, complete structured hook history, source-size architecture guards, Rust 1.85-compatible dependencies, GitHub-hosted Linux/macOS CI, non-locking cancellation, and multimodal image support

Star this repo if you're building AI agents with local models in Rust!