What is webhound-mcp?

Webhound is a tool that gives AI agents a way to budget their research. When you ask an AI to look into something, it often has no clear sense of how much effort to spend. Webhound fixes this by letting you set a dollar limit alongside your prompt. The agent then searches, reads, verifies, and assembles a cited report or structured dataset, stopping when the budget runs out. The final result includes working documents, sources, claim traces, and an evidence pack so you can see how the answer was reached. The project runs as an MCP server, which is a standard protocol that tools like Claude Code, Cursor, GitHub Copilot, and others can speak. You can run it locally on your machine or connect to a hosted version over the internet. Each person authorizes their own Webhound account through OAuth, meaning no shared credentials are bundled into the package. The repository includes client-specific configuration files for several popular coding assistants, making setup straightforward for each one. Installation involves creating a Webhound API key and adding a small configuration block to your agent's settings file. The README provides exact configuration snippets for many clients, including Claude Code, Cursor, Codex, VS Code, Windsurf, and Cline. Some clients support a hosted URL with OAuth, while others use a local setup that runs the tool directly. The package is distributed through npm as webhound-mcp. The research itself is handled by something called Hound, which is the internal system Webhound uses to plan, execute, verify, and assemble results. Hound is built using DeepSeek V4 Pro and GPT-5.4, but it is not a simple pass-through to a single model. The prompt defines the investigation, and the dollar budget controls how much work Hound does before completing. Suggested defaults include a five dollar budget and a report as the output type. The README also covers integration with platforms like Replit, n8n, Smithery, Manus, and ChatGPT's developer mode. Each path follows the same pattern: connect to the hosted URL, authorize your own account, and start a task. The full README is longer than what was shown.