agentool

agentool

21 AI agent tools + context-compaction middleware for the Vercel AI SDK.

File operations, shell execution, code search, web fetching, and more -- everything an AI agent needs to interact with a codebase and system.

Features

• 21 production-ready tools -- bash, grep, glob, read, edit, write, web-fetch, web-search, tool-search, memory, multi-edit, diff, task-create, task-get, task-update, task-list, lsp, http-request, ask-user, sleep
• Context-compaction middleware -- transparent prompt compaction via wrapLanguageModel(), preserves system messages and recent turns
• Vercel AI SDK compatible -- works with generateText(), streamText(), and any AI SDK provider (OpenAI, Anthropic, Google, etc.)
• Factory + default pattern -- createBash({ cwd: '/my/project' }) for custom config, or just use bash with zero config
• Dual ESM/CJS -- works everywhere with proper exports map
• TypeScript-first -- full type declarations, strict mode, no any
• Never throws -- every execute() returns a descriptive error string instead of throwing
• Tree-shakeable -- 22 subpath exports, only import what you need

Installation

npm install agentool ai zod

ai and zod are peer dependencies. You also need an AI SDK provider like @ai-sdk/openai, @ai-sdk/anthropic, etc.

Prerequisites

• Node.js >= 18
• ripgrep (rg) -- required for grep and glob tools

# macOS
brew install ripgrep

# Ubuntu/Debian
sudo apt install ripgrep

# Windows
choco install ripgrep

Quick Start

import { generateText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { bash, read, edit, glob, grep } from 'agentool';

const { text } = await generateText({
  model: openai('gpt-4o'),
  tools: { bash, read, edit, glob, grep },
  maxSteps: 10,
  prompt: 'Find all TypeScript files with TODO comments and list them',
});

Tree-shake with subpath imports

// Only import what you need -- no unused code loaded
import { bash } from 'agentool/bash';
import { grep } from 'agentool/grep';

Custom configuration

Use factory functions when you need to configure tools (custom cwd, timeouts, etc.):

import { createBash } from 'agentool/bash';
import { createRead } from 'agentool/read';

const myBash = createBash({ cwd: '/my/project', timeout: 60000 });
const myRead = createRead({ cwd: '/my/project' });

// Use them like any other tool
const { text } = await generateText({
  model: openai('gpt-4o'),
  tools: { bash: myBash, read: myRead },
  maxSteps: 10,
  prompt: 'List all files and read package.json',
});

Tools Reference

bash

Execute shell commands with timeout and signal handling.

import { bash } from 'agentool/bash';

const result = await bash.execute(
  { command: 'echo hello && ls -la' },
  { toolCallId: 'id', messages: [] },
);

With custom config:

import { createBash } from 'agentool/bash';

const bash = createBash({
  cwd: '/my/project',
  timeout: 60000,    // 60s timeout (default: 120s)
  shell: '/bin/zsh', // custom shell
});

Parameters: command (string), timeout? (number), description? (string)

read

Read files with line numbers, offset/limit pagination, and dual-path reading (fast for <10MB, streaming for larger).

import { read } from 'agentool/read';

// Read entire file
const content = await read.execute(
  { file_path: '/app/src/index.ts' },
  { toolCallId: 'id', messages: [] },
);
// Returns: "1\texport function hello() {\n2\t  return 'world';\n3\t}"

// Read specific range
const range = await read.execute(
  { file_path: '/app/src/index.ts', offset: 10, limit: 20 },
  { toolCallId: 'id', messages: [] },
);

Parameters: file_path (string), offset? (number), limit? (number)

edit

Exact string replacement with curly-quote normalization fallback.

import { edit } from 'agentool/edit';

const result = await edit.execute(
  {
    file_path: '/app/src/config.ts',
    old_string: 'const PORT = 3000;',
    new_string: 'const PORT = 8080;',
  },
  { toolCallId: 'id', messages: [] },
);

// Replace all occurrences
const resultAll = await edit.execute(
  {
    file_path: '/app/src/config.ts',
    old_string: 'localhost',
    new_string: '0.0.0.0',
    replace_all: true,
  },
  { toolCallId: 'id', messages: [] },
);

Parameters: file_path (string), old_string (string), new_string (string), replace_all? (boolean)

write

Write files with automatic parent directory creation.

import { write } from 'agentool/write';

const result = await write.execute(
  {
    file_path: '/app/src/utils/helpers.ts',
    content: 'export function add(a: number, b: number) {\n  return a + b;\n}\n',
  },
  { toolCallId: 'id', messages: [] },
);
// Returns: "Created file: /app/src/utils/helpers.ts (62 bytes)"

Parameters: file_path (string), content (string)

grep

Search file contents with ripgrep. Three output modes, context lines, pagination.

import { grep } from 'agentool/grep';

// Find matching lines with context
const content = await grep.execute(
  { pattern: 'TODO|FIXME', output_mode: 'content', '-C': 2 },
  { toolCallId: 'id', messages: [] },
);

// List files containing matches (sorted by mtime)
const files = await grep.execute(
  { pattern: 'import.*react', output_mode: 'files_with_matches', glob: '*.tsx' },
  { toolCallId: 'id', messages: [] },
);

// Count matches per file
const counts = await grep.execute(
  { pattern: 'console\\.log', output_mode: 'count' },
  { toolCallId: 'id', messages: [] },
);

Parameters: pattern (string), path? (string), output_mode? ('content' | 'files_with_matches' | 'count'), glob? (string), type? (string), -i? (boolean), -n? (boolean), -A? (number), -B? (number), -C? / context? (number), head_limit? (number), offset? (number), multiline? (boolean)

glob

Find files by pattern with ripgrep, sorted by modification time.

import { glob } from 'agentool/glob';

const result = await glob.execute(
  { pattern: '**/*.test.ts' },
  { toolCallId: 'id', messages: [] },
);
// Returns: "Found 27 files\n/app/tests/unit/bash.test.ts\n..."

// Search in specific directory
const result2 = await glob.execute(
  { pattern: '*.json', path: '/app/config' },
  { toolCallId: 'id', messages: [] },
);

Parameters: pattern (string), path? (string)

multi-edit

Atomically apply multiple edits to a single file. All succeed or none are applied.

import { multiEdit } from 'agentool/multi-edit';

const result = await multiEdit.execute(
  {
    file_path: '/app/src/config.ts',
    edits: [
      { old_string: 'const PORT = 3000;', new_string: 'const PORT = 8080;' },
      { old_string: "const HOST = 'localhost';", new_string: "const HOST = '0.0.0.0';" },
    ],
  },
  { toolCallId: 'id', messages: [] },
);

Parameters: file_path (string), edits (array of { old_string, new_string })

diff

Generate unified diffs between files or strings.

import { diff } from 'agentool/diff';

// Compare two files
const fileDiff = await diff.execute(
  { file_path: '/app/old.ts', other_file_path: '/app/new.ts' },
  { toolCallId: 'id', messages: [] },
);

// Compare strings
const stringDiff = await diff.execute(
  { old_content: 'hello world', new_content: 'hello universe' },
  { toolCallId: 'id', messages: [] },
);

Parameters: file_path? (string), other_file_path? (string), old_content? (string), new_content? (string)

web-fetch

Fetch URLs with automatic HTML-to-markdown conversion.

import { webFetch } from 'agentool/web-fetch';

const result = await webFetch.execute(
  { url: 'https://example.com' },
  { toolCallId: 'id', messages: [] },
);
// Returns markdown content (HTML converted via Turndown, truncated at 100K chars)

Parameters: url (string, must be valid URL)

web-search

Search the web with a callback-based implementation (bring your own search provider).

import { createWebSearch } from 'agentool/web-search';

const webSearch = createWebSearch({
  onSearch: async (query, { allowed_domains, blocked_domains }) => {
    // Use Tavily, SerpAPI, Google, or any search provider
    const results = await mySearchProvider.search(query, { allowed_domains, blocked_domains });
    return results.map(r => `${r.title}: ${r.url}\n${r.snippet}`).join('\n\n');
  },
});

const result = await webSearch.execute(
  { query: 'TypeScript best practices 2024', allowed_domains: ['typescript-eslint.io'] },
  { toolCallId: 'id', messages: [] },
);

Parameters: query (string, min 2 chars), allowed_domains? (string[]), blocked_domains? (string[])

tool-search

Search through a registry of available tools by name or keyword.

import { createToolSearch } from 'agentool/tool-search';

const toolSearch = createToolSearch({
  tools: {
    bash: { description: 'Execute shell commands' },
    grep: { description: 'Search file contents with regex' },
    read: { description: 'Read file contents' },
  },
});

const result = await toolSearch.execute(
  { query: 'file', max_results: 3 },
  { toolCallId: 'id', messages: [] },
);
// Returns matching tools sorted by relevance

Parameters: query (string), max_results? (number, default 5)

http-request

Make raw HTTP requests without markdown conversion.

import { httpRequest } from 'agentool/http-request';

const result = await httpRequest.execute(
  {
    method: 'POST',
    url: 'https://api.example.com/data',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ key: 'value' }),
    timeout: 5000,
  },
  { toolCallId: 'id', messages: [] },
);
// Returns JSON: { status, statusText, headers, body }

Parameters: method ('GET' | 'POST' | 'PUT' | 'PATCH' | 'DELETE' | 'HEAD'), url (string), headers? (object), body? (string), timeout? (number)

memory

File-based key-value store for persistent agent memory.

import { memory } from 'agentool/memory';

// Write
await memory.execute(
  { action: 'write', key: 'user-prefs', content: 'Prefers dark mode' },
  { toolCallId: 'id', messages: [] },
);

// Read
const data = await memory.execute(
  { action: 'read', key: 'user-prefs' },
  { toolCallId: 'id', messages: [] },
);

// List all keys
const keys = await memory.execute(
  { action: 'list' },
  { toolCallId: 'id', messages: [] },
);

// Delete
await memory.execute(
  { action: 'delete', key: 'user-prefs' },
  { toolCallId: 'id', messages: [] },
);

Parameters: action ('read' | 'write' | 'list' | 'delete'), key? (string), content? (string)

task-create

Create a new task with subject, description, and optional metadata.

import { taskCreate } from 'agentool/task-create';

const result = await taskCreate.execute(
  { subject: 'Fix login bug', description: 'Auth fails on refresh', metadata: { priority: 'high' } },
  { toolCallId: 'id', messages: [] },
);

Parameters: subject (string), description (string), metadata? (Record<string, unknown>)

task-get

Retrieve a task by ID to see full details.

import { taskGet } from 'agentool/task-get';

const result = await taskGet.execute(
  { taskId: 'abc123' },
  { toolCallId: 'id', messages: [] },
);

Parameters: taskId (string)

task-update

Update a task's status, owner, metadata, and dependency relationships.

import { taskUpdate } from 'agentool/task-update';

// Update status and owner
await taskUpdate.execute(
  { taskId: 'abc123', status: 'in_progress', owner: 'agent-1' },
  { toolCallId: 'id', messages: [] },
);

// Add dependencies and merge metadata (null deletes a key)
await taskUpdate.execute(
  { taskId: 'abc123', addBlockedBy: ['def456'], metadata: { priority: null, notes: 'reviewed' } },
  { toolCallId: 'id', messages: [] },
);

Parameters: taskId (string), subject? (string), description? (string), status? ('pending' | 'in_progress' | 'completed' | 'deleted'), owner? (string), activeForm? (string), addBlocks? (string[]), addBlockedBy? (string[]), metadata? (Record<string, unknown>)

task-list

List all non-deleted tasks with status and dependencies.

import { taskList } from 'agentool/task-list';

const result = await taskList.execute(
  {},
  { toolCallId: 'id', messages: [] },
);

Parameters: none

lsp

Language Server Protocol operations for code intelligence.

import { createLsp } from 'agentool/lsp';

const lsp = createLsp({
  servers: {
    '.ts': { command: 'typescript-language-server', args: ['--stdio'] },
    '.py': { command: 'pylsp' },
  },
});

const result = await lsp.execute(
  { operation: 'goToDefinition', filePath: 'src/index.ts', line: 10, character: 5 },
  { toolCallId: 'id', messages: [] },
);

Parameters: operation ('goToDefinition' | 'findReferences' | 'hover' | 'documentSymbol' | 'workspaceSymbol' | 'goToImplementation' | 'prepareCallHierarchy' | 'incomingCalls' | 'outgoingCalls'), filePath (string), line (number, 1-based), character (number, 1-based)

context-compaction (middleware)

Transparent context compaction middleware for wrapLanguageModel(). When the prompt exceeds a configurable threshold, it automatically summarizes older conversation history while preserving system messages and recent turns.

import { createContextCompaction } from 'agentool/context-compaction';
import { wrapLanguageModel, generateText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';

const model = wrapLanguageModel({
  model: anthropic('claude-sonnet-4-20250514'),
  middleware: createContextCompaction({
    maxContextTokens: 200_000,        // model's context window (required)
    autoCompactThresholdPct: 0.80,    // compact when 80% full (default)
    summaryTargetPct: 0.05,           // summarize to 5% of context (default)
  }),
});

// Use the wrapped model normally — compaction is transparent
const { text } = await generateText({
  model,
  tools: { bash, read, edit },
  maxSteps: 20,
  prompt: 'Find and fix the bug in src/auth.ts',
});

Config: maxContextTokens (number, required), autoCompactThresholdPct? (0-1), summaryTargetPct? (0-1), reservedOutputTokens? (number), estimateTokens? (function), summarize? (function), onCompactionFailure? ('passthrough' | 'throw')

ask-user

Prompt the user for input during agent execution.

import { createAskUser } from 'agentool/ask-user';

const askUser = createAskUser({
  onQuestion: async (question, options) => {
    // Your UI logic to prompt the user
    return 'User response here';
  },
});

const answer = await askUser.execute(
  { question: 'Which database should I use?', options: ['PostgreSQL', 'MySQL', 'SQLite'] },
  { toolCallId: 'id', messages: [] },
);

Parameters: question (string), options? (string[])

sleep

Pause execution for rate limiting or polling intervals.

import { sleep } from 'agentool/sleep';

const result = await sleep.execute(
  { durationMs: 2000, reason: 'Waiting for deployment' },
  { toolCallId: 'id', messages: [] },
);
// Returns: "Slept for 2001ms. Reason: Waiting for deployment"

Parameters: durationMs (number, max 300000), reason? (string)

Configuration

Every tool follows the factory + default pattern:

// Default instance -- uses process.cwd(), default timeouts
import { bash } from 'agentool/bash';

// Custom instance -- configure cwd, timeouts, and tool-specific options
import { createBash } from 'agentool/bash';
const myBash = createBash({
  cwd: '/my/project',
  timeout: 60000,
  shell: '/bin/zsh',
});

Base configuration

All tools accept BaseToolConfig:

Option	Type	Default	Description
cwd	string	process.cwd()	Working directory for file operations

Tools that support timeouts extend TimeoutConfig:

Option	Type	Default	Description
timeout	number	varies	Timeout in milliseconds

Tool-specific configuration

Tool	Extra Config
bash	shell?: string -- shell binary path
read	maxLines?: number -- max lines to return (default: 2000)
memory	memoryDir?: string -- storage directory
task-create	tasksFile?: string -- JSON file path
task-get	tasksFile?: string -- JSON file path
task-update	tasksFile?: string -- JSON file path
task-list	tasksFile?: string -- JSON file path
web-search	onSearch?: (query, opts) => Promise<string> -- search callback
tool-search	tools?: Record<string, { description }> -- tool registry
lsp	servers?: Record<string, LspServerConfig> -- LSP servers by file extension
http-request	defaultHeaders?: Record<string, string> -- headers merged into every request
web-fetch	maxContentLength?: number, userAgent?: string
ask-user	onQuestion?: (question, options?) => Promise<string>
sleep	maxDuration?: number -- cap in ms (default: 300000)

Error Handling

Every tool's execute() catches errors internally and returns a descriptive string -- it never throws:

const result = await read.execute(
  { file_path: '/nonexistent/file.ts' },
  { toolCallId: 'id', messages: [] },
);
// Returns: "Error [read]: Failed to read file: ENOENT: no such file or directory..."

Error strings follow the format: Error [tool-name]: {description} with actionable context to help the AI model recover.

Imports

Barrel import (all tools)

import {
  bash, createBash,
  read, createRead,
  edit, createEdit,
  write, createWrite,
  grep, createGrep,
  glob, createGlob,
  webFetch, createWebFetch,
  webSearch, createWebSearch,
  toolSearch, createToolSearch,
  httpRequest, createHttpRequest,
  memory, createMemory,
  multiEdit, createMultiEdit,
  diff, createDiff,
  taskCreate, createTaskCreate,
  taskGet, createTaskGet,
  taskUpdate, createTaskUpdate,
  taskList, createTaskList,
  lsp, createLsp,
  createContextCompaction,  // middleware, not a tool
  askUser, createAskUser,
  sleep, createSleep,
} from 'agentool';

Subpath imports (tree-shakeable)

import { bash } from 'agentool/bash';
import { grep } from 'agentool/grep';
import { glob } from 'agentool/glob';
import { read } from 'agentool/read';
import { edit } from 'agentool/edit';
import { write } from 'agentool/write';
import { webFetch } from 'agentool/web-fetch';
import { httpRequest } from 'agentool/http-request';
import { memory } from 'agentool/memory';
import { multiEdit } from 'agentool/multi-edit';
import { diff } from 'agentool/diff';
import { taskCreate } from 'agentool/task-create';
import { taskGet } from 'agentool/task-get';
import { taskUpdate } from 'agentool/task-update';
import { taskList } from 'agentool/task-list';
import { webSearch } from 'agentool/web-search';
import { toolSearch } from 'agentool/tool-search';
import { lsp } from 'agentool/lsp';
import { createContextCompaction } from 'agentool/context-compaction'; // middleware
import { askUser } from 'agentool/ask-user';
import { sleep } from 'agentool/sleep';

Full Example: AI Coding Agent

import { generateText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { bash, read, edit, write, glob, grep, diff } from 'agentool';

const { text, steps } = await generateText({
  model: openai('gpt-4o'),
  tools: { bash, read, edit, write, glob, grep, diff },
  maxSteps: 20,
  system: `You are a coding assistant. You can read, search, edit, and write files.
    Always read a file before editing it. Use grep to search for patterns.
    Use glob to find files. Use bash for git, build, and test commands.`,
  prompt: 'Find all console.log statements in src/ and replace them with proper logger calls',
});

console.log(`Completed in ${steps.length} steps`);
console.log(text);

Requirements

Dependency	Version	Required
Node.js	>= 18	Yes
ai (Vercel AI SDK)	>= 5.0.17	Peer dependency
zod	>= 3.23.0	Peer dependency
ripgrep (rg)	any	For grep/glob tools

License

Apache-2.0