cnctd_ai

A Rust abstraction layer for AI/LLM providers (Anthropic Claude, OpenAI, Google Gemini, OpenRouter) with integrated MCP (Model Context Protocol) support and autonomous agent framework.

Features

• Multi-Provider Support: Unified interface for Anthropic Claude, OpenAI, Google Gemini, and OpenRouter
• Streaming & Non-Streaming: Support for both regular completions and streaming responses
• Tool Calling: Full support for function/tool calling across all providers
• Agent Framework: Autonomous task execution with tool calling loops
• MCP Integration: Native support for MCP servers (stdio and HTTP gateway)
• Error Handling: Comprehensive error types with provider-specific handling
• Type Safety: Strong typing with proper error handling throughout

Installation

cargo add cnctd_ai

Quick Start

Agent Framework

The easiest way to build autonomous AI applications:

use cnctd_ai::{Agent, Client, AnthropicConfig, McpGateway};

// Setup client and gateway
let client = Client::anthropic(
    AnthropicConfig {
        api_key: "your-key".into(),
        model: "claude-sonnet-4-20250514".into(),
        version: None,
    },
    None,
)?;

let gateway = McpGateway::new("https://mcp.cnctd.world");

// Create agent with default settings
let agent = Agent::new(&client).with_gateway(&gateway);

// Run autonomous task - agent will use tools as needed
let trace = agent.run_simple(
    "Research the latest Rust async trends and summarize key findings"
).await?;

// View results
trace.print_summary();

For advanced configuration:

let agent = Agent::builder(&client)
    .max_iterations(10)
    .max_duration(Duration::from_secs(300))
    .system_prompt("You are a helpful research assistant.")
    .gateway(&gateway)
    .build();

See Agent Framework Documentation for more details.

Basic Completion

use cnctd_ai::{Client, AnthropicConfig, Message, CompletionRequest};

let client = Client::anthropic(
    AnthropicConfig {
        api_key: "your-api-key".into(),
        model: "claude-sonnet-4-20250514".into(),
        version: None,
    },
    None,
)?;

let request = CompletionRequest {
    messages: vec![Message::user("Hello, how are you?")],
    tools: None,
    options: None,
};

let response = client.complete(request).await?;
println!("Response: {}", response.text());

Streaming

use cnctd_ai::{Client, AnthropicConfig, Message, CompletionRequest};

let mut stream = client.complete_stream(request).await?;

while let Some(chunk) = stream.next().await {
    let chunk = chunk?;
    if let Some(text) = chunk.text() {
        print!("{}", text);
    }
}

Tool Calling

use cnctd_ai::{Client, Message, CompletionRequest, create_tool};
use serde_json::json;

// Create a tool using the helper function
let weather_tool = create_tool(
    "get_weather",
    "Get the current weather for a location",
    json!({
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "The city and state, e.g. San Francisco, CA"
            },
            "unit": {
                "type": "string",
                "enum": ["celsius", "fahrenheit"]
            }
        },
        "required": ["location"]
    })
)?;

let mut request = CompletionRequest {
    messages: vec![Message::user("What's the weather in SF?")],
    tools: None,
    options: None,
};
request.add_tool(weather_tool);

let response = client.complete(request).await?;

// Check if model wants to use a tool
if let Some(tool_use) = response.tool_use() {
    println!("Tool: {}", tool_use.name);
    println!("Arguments: {}", tool_use.input);
    
    // Execute tool and continue conversation
    // See examples/tool_calling.rs for full implementation
}

MCP Gateway Integration

use cnctd_ai::McpGateway;

let gateway = McpGateway::new("https://mcp.cnctd.world");

// List available servers
let servers = gateway.list_servers().await?;

// List tools from a specific server
let tools = gateway.list_tools("brave-search").await?;

// Execute a tool
let result = gateway.call_tool(
    "brave-search",
    "brave_web_search",
    Some(json!({"query": "Rust programming"})),
).await?;

Examples

The repository includes several examples:

Agent Framework:

• agent_simple.rs - Minimal agent setup
• agent_basic.rs - Full-featured agent with configuration

Core Functionality:

• basic_completion.rs - Simple completion example
• streaming.rs - Streaming responses
• tool_calling.rs - Function/tool calling
• tool_calling_streaming.rs - Tool calling with streaming
• conversation.rs - Multi-turn conversations
• error_handling.rs - Error handling patterns
• mcp_gateway.rs - MCP gateway integration

Run examples with:

cargo run --example agent_simple
cargo run --example basic_completion

Environment Variables

Set these for the examples:

ANTHROPIC_API_KEY=your-anthropic-key
OPENAI_API_KEY=your-openai-key
GEMINI_API_KEY=your-gemini-key
GATEWAY_URL=https://mcp.cnctd.world  # Optional
GATEWAY_TOKEN=your-token  # Optional

Tool Creation Helpers

The library provides helper functions for easier tool creation:

use cnctd_ai::{create_tool, create_tool_borrowed};

// For owned strings (runtime data)
let tool = create_tool(name, description, schema)?;

// For static strings (compile-time constants)
let tool = create_tool_borrowed(name, description, schema)?;

OpenAI Responses API & Multi-Turn Tool Calls

cnctd_ai uses OpenAI's newer Responses API (/v1/responses) for GPT-4, GPT-4.1, GPT-5, and reasoning models (o1, o3). This provides better tool calling support but has specific requirements for multi-turn conversations:

Key Concepts

• call_id: OpenAI uses call_id (format: call_...) to match function_call items with their function_call_output responses
• ToolUse.call_id: The library captures this from API responses and stores it in ToolUse.call_id
• ToolResult.effective_call_id(): Returns the correct ID to use when sending tool results back

Reasoning Models (GPT-5.2-pro, o1, o3)

Reasoning models require special handling for multi-turn tool calls:

1. Encrypted reasoning content: The library automatically requests reasoning.encrypted_content for reasoning models
2. Reasoning items: Must be echoed back in continuation requests via Message.reasoning_items

The library handles this automatically - just ensure you preserve reasoning_items when building continuation messages:

// After getting a response with tool calls
let response = client.complete(request).await?;

// The response.message includes reasoning_items if present
// When building the next request, include the full message
messages.push(response.message.clone());

// Add tool results
let tool_result = ToolResult::new(tool_use.call_id.unwrap_or(tool_use.id), output);
messages.push(Message::tool_results(vec![tool_result]));

Application Considerations

When persisting and reconstructing conversations from a database:

1. Store call_id: Save both tool_use_id and call_id from tool calls
2. Match 1:1: Every function_call must have a matching function_call_output
3. Preserve reasoning: Store and restore reasoning_items for reasoning models

Error Handling

The library provides comprehensive error types:

use cnctd_ai::Error;

match client.complete(request).await {
    Ok(response) => { /* handle success */ },
    Err(Error::AuthenticationFailed(msg)) => { /* handle auth */ },
    Err(Error::RateLimited { retry_after }) => { /* handle rate limit */ },
    Err(Error::ProviderError { provider, message, status_code }) => { /* handle provider error */ },
    Err(e) => { /* handle other errors */ },
}

cnctd_ai_server

The repository includes a subcrate at crates/cnctd_ai_server/ -- an Axum-based REST API that provides:

• Streaming SSE chat with full tool-calling loops
• MCP integration for tool discovery and execution
• 4-point data obfuscation protecting sensitive entity data during AI conversations
• Agent execution with background task management

The obfuscation system is fully dynamic -- the server fetches its entity dictionary from an HTTP source URL hosted by your application. No entity types are hardcoded.

See the Obfuscation Setup Guide for integration details.

Documentation

• Agent Framework Guide - Complete agent framework documentation
• Obfuscation Setup Guide - Integration guide for cnctd_ai_server obfuscation
• API Documentation - Full API reference

License

MIT License - see LICENSE file for details.