agentool

Lightweight, provider-agnostic AI agent toolkit for CI/CD. Built-in bash, edit, grep, git, and HTTP tools with modular add-ons for Docker, test runners, and MCP.

What it is

A single npm package that gives AI agents a complete set of system tools for CI/CD automation. It wraps Vercel AI SDK v6 (ai package) and provides prebuilt tools, a simple agent runner, and modular add-ons via Node.js subpath exports.

The problem: if you want an AI agent that can run bash, edit files, grep code, call APIs, and interact with GitHub in a CI pipeline, you either use Claude Code (locked to Anthropic) or wire up a dozen libraries yourself. agentool gives you all of that out of the box, with any LLM provider — OpenAI, Anthropic, LiteLLM, or any OpenAI-compatible endpoint.

Tech stack

• TypeScript (strict), Node.js 20+
• Build: tsup
• AI SDK: Vercel AI SDK v6 (ai) as peer dependency
• Testing: vitest
• Package manager: pnpm
• License: Apache-2.0

Package structure

Single npm package, multiple entry points via subpath exports. Users install once (npm install agentool) and import only what they need.

import { createAgent, coreTools } from "agentool/core";
import { dockerTools } from "agentool/docker";
import { testRunnerTools } from "agentool/test-runner";

Modules

agentool/core (must-have, zero optional deps)

• createAgent — agent runner wrapping AI SDK ToolLoopAgent, auto-assembles system prompt from tool descriptions
• bash — shell command execution via child_process
• read_file — read file contents with optional line range
• write_file — create/overwrite files
• edit — str_replace style find-and-replace (oldStr must be unique in file)
• grep — structured search, uses ripgrep if available, falls back to native grep
• find — find files by glob pattern
• git — diff, log, status, blame, show
• gh — GitHub CLI wrapper (pr, issue, workflow, api)
• http — structured HTTP requests via native fetch

agentool/docker (add-on)

• docker build, run, compose, logs

agentool/test-runner (add-on)

• Generic test runner with structured output parsing (jest, vitest, pytest)

agentool/data-parse (add-on)

• jq, yq, csv parsing

agentool/code-intel (add-on)

• ast-grep, directory tree, symbol extraction

agentool/mcp (add-on)

• Generic MCP client bridge — connect to any MCP server and expose its tools

Usage

import { createAgent, coreTools } from "agentool/core";

const agent = createAgent({
  model: "openai/gpt-4o",
  // Or LiteLLM: baseURL: "http://localhost:4000/v1"
  tools: { ...coreTools() },
});

const result = await agent.run("Find why the tests are failing and fix it");

Implementation priorities

1. Core types (ToolDefinition, ToolResult, ToolPack interfaces)
2. Core tools (bash, read_file, write_file, edit, grep, find, git, gh, http)
3. createAgent runner
4. System prompt auto-assembly
5. Add-on modules (docker, test-runner, data-parse, code-intel, mcp)
6. Tests for each tool