GitHub - CharlyCst/spadebox: Sandboxed tools and JS runtime for AI agents

Sandboxed tools and JS runtime for your AI agents

SpadeBox is a set of sandboxed tools and a JS runtime for AI agents, written in Rust with JavaScript and Python bindings. Focus on your domain-specific tools and harness — give your agent SpadeBox for the rest.

Category	Tools
Files	`read_file`, `write_file`, `edit_file`, `move`, `grep`, `glob`
Network	`fetch`
Code execution	`js_repl`, `js_exec`

Features

Lightweight sandboxing: SpadeBox uses the cap-std crate for file system sandboxing, and domain allowlisting for HTTP requests. The JS engine is based on boa, and uses the same sandboxing policies.
Configurable: Pick only the tools you need for your application: files, network, or code execution, or any combination of those.
No bash tool: SpadeBox has been designed to work well even without a bash tool. The tools are designed to cover all the basic operations an agent needs without having to shell-out. If you would like your agent to use bash you can provide your own bash tool in addition to SpadeBox.
Native function in JS runtime: Expose native functions to the SpadeBox JS runtime to allow your agents to programmatically interact with your application. Compatible with all supported host language bindings.
Secret management for HTTP requests: Register credentials for specific HTTP domains and get a token that can be safely shared with your agent. Spadebox replaces the token by the actual secret within HTTP requests to the target domain.
Default limits to preserve context: SpadeBox's tools try to safeguard your agent context, with default limits to tool outputs and HTML-to-markdown conversion.

Usage

(checkout out the documentation for more)

JavaScript

import { SpadeBox } from "@spadebox/spadebox";

const sb = new SpadeBox()
  .enableFiles("/workspace")
  .enableHttp()
  .allow("api.example.com", ["GET", "POST"])
  .enableJs();

const tools = sb.tools(); // pass to your LLM as available tools

// dispatch a tool call coming from the model
const result = await sb.callTool("read_file", JSON.stringify({ path: "src/main.rs" }));

Rust

use spadebox_core::{Sandbox, DomainRule, HttpVerb, enabled_tools, call_tool};

let mut sandbox = Sandbox::new();
sandbox
    .enable_fs("/workspace")?
    .enable_http()
    .allow(DomainRule::new("api.example.com", vec![HttpVerb::Get, HttpVerb::Post])?)
    .enable_js();

let tools = enabled_tools(&sandbox); // pass to your LLM as available tools

// dispatch a tool call coming from the model
let result = call_tool(&sandbox, "read_file", r#"{"path":"src/main.rs"}"#.into()).await?;

Python

from spadebox import SpadeBox

sb = (SpadeBox()
    .enable_files("/workspace")
    .enable_http()
    .allow("api.example.com", ["GET", "POST"])
    .enable_js())

tools = sb.tools()  # pass to your LLM as available tools

# dispatch a tool call coming from the model
result = sb.call_tool("read_file", '{"path": "src/main.rs"}')

MCP

# filesystem tools only
spadebox-mcp --files /workspace

# HTTP tools only (allow specific domains and verbs)
spadebox-mcp --allow "api.example.com:GET,POST" --allow "*.cdn.example.com:GET"

# JavaScript REPL only
spadebox-mcp --js

# all tools
spadebox-mcp --files /workspace --allow "api.example.com:GET" --js

推荐订阅源

Show HN