The unedited session: list_providers, search, search_rag, and a cited answer, all in the terminal. With narration.
Agents are only as useful as what they can reach. Give a model no tools and it answers from memory; give it real retrieval and it starts behaving like a colleague with access. This demo is a recording of exactly that transition: a stock Claude CLI, three SWIRL tools over MCP, one prompt, no editing.
The session, play by play
First, Claude calls list_providers and learns what this deployment can see: 32 sources, from Microsoft 365 to databases to the open web, returned as structured data rather than marketing copy.
Next, a single search call federates the query across all of them. The response carries 288 million total hits, ranked, and each result arrives with title, snippet, author, date, source, and relevance scores, everything an agent needs to reason about what it found.
Finally, search_rag does the grounding: a synthesized answer with citations, including deep links that open the source document scrolled to the quoted passage.
The moment nobody scripted
Along the way, Claude noticed something we did not ask it to look for: three distinct versions of the policy were in circulation across two systems, and it said so. That is the practical argument for giving agents real retrieval instead of a curated snapshot. An agent that can see the enterprise as it actually is will surface the problems a snapshot would have hidden.
Built in, permissions included
SWIRL 5 ships its MCP server as part of the product. Any MCP client, Claude Desktop, Claude Code, Cursor, or your own agent framework, gets federated search, document reading, and grounded RAG as native tools. The agent operates under a defined identity, so what it can see is exactly what that identity can see; permissions are enforced at the source on every call.
Setup takes minutes with the MCP Guide, or let the SWIRL Plugin for Claude Code wire it up for you with /swirl:mcp. Building an agent that needs enterprise reach? Talk to us.