Crate agent_code_lib 

agent-code-lib

The complete AI coding agent engine as a reusable library.

This crate contains everything needed to build an AI coding agent: LLM providers, tools, the query engine, memory, permissions, and all supporting services. The agent CLI binary is a thin wrapper over this library.

Modules

Module	Purpose
config	Configuration loading and schema (TOML, layered merge)
error	Unified error types (LlmError, ToolError, ConfigError)
hooks	Lifecycle hooks (pre/post tool use, session events)
llm	LLM communication: streaming client, message types, providers, retry
memory	Persistent context: project (AGENTS.md) and user memory
permissions	Permission system: rules, modes, protected directories
query	Agent loop: call LLM → execute tools → compact → repeat
services	Cross-cutting: tokens, compaction, sessions, git, MCP, plugins, diagnostics
skills	Custom workflow loading from markdown files
state	Session state: messages, usage tracking, cost
tools	32 built-in tools and the Tool trait for custom tools

Quick Example

use agent_code_lib::config::Config;
use agent_code_lib::tools::registry::ToolRegistry;
use agent_code_lib::state::AppState;

let config = Config::default();
let tools = ToolRegistry::default_tools();
let state = AppState::new(config);
// state.messages, state.total_cost_usd, etc. are now available

Adding a Custom Tool

Implement the tools::Tool trait:

use async_trait::async_trait;
use agent_code_lib::tools::{Tool, ToolContext, ToolResult};

struct MyTool;

#[async_trait]
impl Tool for MyTool {
    fn name(&self) -> &'static str { "MyTool" }
    fn description(&self) -> &'static str { "Does something useful" }
    fn input_schema(&self) -> serde_json::Value {
        serde_json::json!({"type": "object", "properties": {}})
    }
    fn is_read_only(&self) -> bool { true }
    async fn call(
        &self,
        input: serde_json::Value,
        ctx: &ToolContext,
    ) -> Result<ToolResult, agent_code_lib::error::ToolError> {
        Ok(ToolResult { content: "done".into(), is_error: false })
    }
}

Then register it: registry.register(Arc::new(MyTool));

Modules

config

Configuration system.

error

Unified error types for the application.

hooks

Hook system.

llm

LLM client layer.

memory

Memory system — 3-layer architecture.

permissions

Permission system.

query

Query engine: the core agent loop.

sandbox

Process-level sandboxing for subprocess-spawning tools.

schedule

Scheduled agent execution.

services

Core services layer.

skills

Skill system.

state

Application state management.

tools

Tool system.