Agent Tools

中文 | English

A Rust toolkit for AI agents: workspace file I/O, search, web fetch, Markdown helpers, Git, cross-session memory, pluggable human-in-the-loop I/O, and persistent todos. Each tool is described with JSON Schema and lines up with OpenAI Function Calling, Anthropic Tool Use, and similar formats.

Quick start

Enable only what you need in Cargo.toml:

[dependencies]
agentool = { version = "0.1", features = ["fs", "find", "web", "git"] }

All published features:

agentool = { version = "0.1", features = ["full"] }

Heavy crates (tokio, reqwest, git2, walkdir, regex, chrono, …) are optional: each feature in Cargo.toml pulls only what it needs. The always-on footprint is async-trait, serde_json, and thiserror (plus the Tool trait when you enable no features).

Example: write and read a file

use std::sync::Arc;

use agentool::Tool;
use agentool::fs::{all_tools, FsContext};

async fn example() -> Result<(), agentool::ToolError> {
    let ctx = Arc::new(FsContext::new(None, false).expect("workspace root"));
    let tools = all_tools(ctx);

    let write = tools.iter().find(|t| t.name() == "write_file").unwrap();
    let read  = tools.iter().find(|t| t.name() == "read_file").unwrap();

    write.execute(serde_json::json!({
        "path": "example.txt",
        "content": "hello\n",
    })).await?;

    let out = read.execute(serde_json::json!({ "path": "example.txt" })).await?;
    println!("{}", out["data"]["content"]);
    Ok(())
}

Features

exec / code / office / browser / design / gui are planned but not shipped yet.

Tool reference

Each feature has a README.md (English) next to mod.rs under src/<feature>/. Chinese versions live in README.zh.md in the same folder.

Placeholder modules: src/exec/README.md · src/code/README.md · src/office/README.md · src/browser/README.md · src/design/README.md · src/gui/README.md (each has a README.zh.md).

Documentation hub

• docs/README.md — how docs are organized, conventions, links
• docs/README.zh.md — 中文索引

Response shape

All tools return JSON in a single envelope.

Success

{
  "success": true,
  "data": {}
}

Failure

{
  "success": false,
  "error": {
    "code": "ERROR_CODE",
    "message": "Human-readable message"
  }
}

In Rust, Tool::execute returns Result<serde_json::Value, ToolError>: on success the Ok value already includes the success / data shell; hosts map Err(ToolError) to success: false.

License

MIT