MCP-Native Tools for Real-World Access

Web Fetch

Retrieve a web page and convert its content into clean, structured Markdown

Web Fetch combines a high-performance HTTP client with sophisticated HTML-to-Markdown transformation. It handles modern web protocols, redirects, and compression while preserving document structure and semantics.

The tool navigates JavaScript-rendered pages, preserves complex table structures, maintains heading hierarchies, and intelligently handles inline styles. With millisecond processing speeds and smart content filtering, it serves as the critical bridge between HTML content and the structured text that AI agents can efficiently parse and understand.

Common Use Cases: Research and fact-checking, documentation extraction and archiving, automated content synthesis, knowledge base construction

Web Search

Search the web for relevant information on any given topic

Web Search provides AI agents with real-time access to current web information through intelligent, context-aware search capabilities. Unlike traditional search engines that return raw links, this tool delivers pre-processed content snippets optimized for AI consumption.

It understands query context and returns focused results from authoritative sources, filtering out irrelevant content while preserving key information. The service returns structured data including relevant text snippets, source URLs, and publish dates, enabling agents to quickly assess reliability and synthesize findings.

Search Modes: Standard (quick results for factual queries), Deep (comprehensive analysis from multiple sources)

Perfect For: Fact-checking and verification, real-time information retrieval, research beyond training data cutoffs, current events monitoring

HTTP Client

Send HTTP requests (GET, POST, etc.) and receive the raw response, including status, headers, and body

The HTTP Client provides AI agents with direct access to web APIs and HTTP endpoints through a simple yet powerful interface. Built on a high-performance Rust foundation, it supports all standard HTTP methods with full control over request construction.

Agents can set custom headers for authentication and content negotiation, send JSON or form-encoded request bodies, and handle various response formats. With configurable timeout settings and automatic redirect following, it adapts to different API requirements and network conditions.

Built for AI Workflows: Authenticate with APIs using bearer tokens or API keys, submit data to webhooks and trigger automations, retrieve structured data from REST endpoints, chain multiple API calls for complex integrations

Code Runner

Execute Python and JavaScript code instantly in secure WebAssembly sandboxes with reliable performance

Code Runner delivers secure code execution through WebAssembly isolation, creating a fresh sandbox for every request. Each execution runs in a new WASM instance with isolated memory, ensuring complete security between requests.

With instantiation overhead under 10ms, you get both strong isolation and excellent performance. The service includes Python standard library and modern JavaScript (ES6+), enabling common operations like data transformations, calculations, and string manipulation without external dependencies.

Best For: Data transformations and JSON processing, mathematical calculations and algorithms, string manipulation and regex operations, template rendering and text processing

Virtual Machine

Configurable Linux containers (0.5-8 vCPUs) for data analysis, web scraping, and complex computations

Virtual Machine provides full-featured Linux environments where agents can install any Python, Node.js, or system packages - from pandas and numpy to TensorFlow and beyond. Choose from five sizes (XS to XL) to match your workload. Each session gets a dedicated container with filesystem integration, allowing seamless access to uploaded files and persistent result storage.

Complete OS-level access enables running shell commands, managing processes, and interacting with the full Linux environment. Scale from lightweight scripts (0.5 vCPU) to ML workflows (8 vCPUs) - perfect for workflows that need real computing power with transparent per-second pricing.

Perfect For: Data science and machine learning workflows, web scraping with Selenium or Playwright, image and video processing, package installation (pip, npm, apt-get)

Knowledge Retrieval

Search through indexed knowledge bases and documents

Knowledge Retrieval brings RAG capabilities to your agents through hybrid search that combines keyword matching with deep semantic understanding. With sub-100ms latency, it's optimized for real-time agent interactions that demand instant access to information. The system automatically selects and formats the most relevant chunks for LLM consumption, ensuring optimal context usage. Transparent relevance scoring with explainable rankings makes it easy to debug and optimize retrieval strategies, giving you full visibility into why specific results were selected.

Applications: RAG-powered question answering, document search and analysis, knowledge base queries, contextual information retrieval

Files

Session-scoped storage with intelligent lifecycle management across working, input, and output scopes

Files provides AI agents with unified storage across three scopes: working for agent workspace, input for user-provided data, and output for persistent results. Each scope has tailored permissions and automatic cleanup policies that balance speed, cost, and data retention.

Unlike single-purpose storage, Files understands your agent's workflow. Temporary processing in working is fast and auto-cleaned (48 hours), while input and output provide persistent storage for data that matters. Agents read user data from input, work in the working directory, and write results to output - creating a natural dataflow from source to destination.

Perfect For: Multi-step data processing with temporary intermediate files, VM-based code execution with mounted filesystem access, workflows with clear input→process→output dataflow, projects where most data is temporary but results are permanent