Agents are only as good as what they can read
An agent working in your codebase or your support queue will make decisions based on the context it has. When the useful knowledge lives in design notes, runbooks, or policy documents, the agent either searches it or guesses.
Retrieval closes that gap. Instead of pasting documents into a prompt, the agent asks a question and gets back relevant passages with source references, so its answer can point to where the information came from.
Connecting is a prompt, not a project
WonderSearch supports the Model Context Protocol, the standard many agents use to reach outside tools. From the console, you can copy a setup prompt for Codex, Claude Code, Cursor, or another MCP client and paste it into the agent.
The agent then asks you to approve the connection in your browser. That is where you choose which workspaces, drives, and actions the client may use, and you can review or revoke connected clients from the console later.
If you prefer to wire search into your own application instead, the same search is available through the REST API and the Python SDK.
Keep the agent honest
Ask the agent to cite the passages it used. Read a few of them yourself, especially early on. If an answer leans on a source that does not quite say what the agent claims, that is a retrieval question worth investigating, not just a prompt to rewrite.
Grant the smallest access that does the job, and review connected clients from time to time. A focused collection with clear permissions tends to give an agent better context than access to everything.