WAY 1
A copy inside every agent
- CRM agent copy #1 fresh
- Slack bot copy #2 aging
- Service desk copy #3 stale
Load the knowledge into every agent separately and end up with several copies of the same base, each at a different stage of staleness.
MCP server
Your company knowledge as a tool for your agents
Ragen exposes an MCP server. It is a gateway your agents query for company knowledge, and the answer comes back with citations, inside the permissions of the person whose API key they use.
Ragen as an MCP gateway in the middle, connected to Claude Desktop, Cursor, a CRM agent, a Slack bot, a service desk agent, your own code and systems that do not exist yet.
The problem
An agent with no access to company knowledge guesses
Agents are good at doing things. They send the email, open the task, update the CRM record. Knowledge is where it falls apart. Procedures, decisions, contracts and documentation sit in a dozen places, and an agent reaches them only when somebody pastes a fragment into the prompt by hand.
There are two ways out.
WAY 2
One knowledge layer for all of them
Stand up one knowledge layer that all of them use.
Ragen is the second one.
One knowledge base, many agents
From your documents to an agent's answer in three steps.
- 01
Knowledge goes into Ragen
Documents, drives, systems. Ragen indexes them, enforces permissions at the level of a single document, and keeps all of it on your own infrastructure.
- 02
Ragen opens the gateway
An AI Assistant built in Ragen becomes a tool that any MCP client with a Ragen API key can call.
- 03
Your agents ask
An agent in Claude Desktop, in Cursor, in your own code, or in a system that does not exist yet. The answer comes back with citations, and only from the documents that person is allowed to see.
An agent gets no more than a person would
Access control works at document level and holds whether the question comes from a person in a browser or from an agent through the gateway. Every answer comes back with citations, so you can check where the agent got what it wrote.
Without that, an agent stays in a sandbox, well away from company knowledge.
What this makes possible
The same knowledge, wherever the work happens.
No switch to yet another chat window. The answer arrives in the tool people already use, with its source.
CRM
An assistant inside your CRM answers from your own material
A salesperson asks about warranty terms for a customer in food manufacturing. The agent in the CRM pulls the answer from Ragen, together with the contract it came from.
SLACK
Your Slack bot stops pointing people at the documentation
Instead of a link to a folder you get the answer and its source. The bot keeps no copy of the knowledge, it asks Ragen.
SERVICE DESK
A service desk agent sees what the requester sees
Permissions travel with the question. An agent using the API key of someone on the production floor will not see board documents.
CLAUDE DESKTOP · CURSOR
Your team works with company knowledge in the tools it already uses
Claude Desktop, Cursor and other MCP clients, with no switch to yet another chat window.
For technical teams
You write it once, it works with every client
MCP is an open protocol. The next tool that shows up in the company connects to the same gateway, with no new integration to write and no per-tool plugin to maintain.
The gateway runs one way: Ragen hands knowledge out. Connectors for Google Drive, HubSpot or ClickUp are a separate matter, described under Integrations. That is where Ragen pulls knowledge in. See the integrations
Three routes into the same knowledge layer
MCP server
The Ragen AI Assistant as a tool for external MCP clients, Claude Desktop and Cursor among them.
OpenAI-compatible API
If your code talks to OpenAI, it talks to Ragen.
TypeScript SDK
The official @webamigos/ragen-sdk-ts package, the same one we use ourselves.
The tools the gateway exposes
- ragen_chat
- Send a question and get a finished answer back. The API key decides which assistant answers, or you name one.
- ragen_search_knowledge_base
- Search an assistant's knowledge base and get the most relevant document fragments back, with no answer generated.
- ragen_list_assistants
- List the assistants available to your API key, to find the id to call.
claude_desktop_config.json
In Claude Desktop, put the mcp-remote bridge in front of this address. claude_desktop_config.json starts a local process and has no field for a URL or headers, so mcp-remote forwards the calls to Ragen with the Authorization header. The full entry is in the MCP server docs.
The block below is the Cursor entry, the .cursor/mcp.json file. The address is your own Ragen instance, and you authenticate with the same API key as the REST API.
.cursor/mcp.json
{
"mcpServers": {
"ragen": {
"url": "https://your-ragen-instance.example.com/mcp",
"headers": {
"Authorization": "Bearer YOUR_API_KEY"
}
}
}
} Full MCP server documentation
- Hybrid search (dense and BM25)
- Cross-encoder reranking
- Citations
- Multi-tenancy
- Conversation encryption
Next.js · NestJS · BullMQ on Redis or Temporal.io · Postgres · Qdrant · Docling
your infrastructure · open source licence
Connect your agents to your company knowledge
We will show you what it looks like on our side: the Ragen AI Assistant called from Claude Desktop, with permissions and citations.