A RAG knowledge base your AI tools reach over MCP.
RAG as a service for your team's docs and runbooks. CoFix chunks, embeds and links every page, and serves it over MCP to Claude, Claude Code, ChatGPT, GitHub Copilot and Gemini CLI. Retrieval uses graph and vector search and stays within each member's access. There is no pipeline, vector database or MCP server to run.
How it works
From a page your team writes to an answer in their AI tool.
Write
Your team keeps its docs and runbooks in knowledge bases. New pages that fail the profile check wait in the review queue.
Index
The app chunks and embeds each page and stores the links between pages as a graph. There is no pipeline to set up.
Connect you
Each member adds their own MCP connector to Claude, Claude Code, ChatGPT, GitHub Copilot or Gemini CLI, signed in as them.
Retrieve
As the AI tool works, it searches over MCP with graph and vector search, and gets back only what that member can see.
Answer
It answers from your documents, tickets and work log, and every action it takes is in the activity log.
RAG vs MCP
MCP connects. RAG finds the answer.
The two are often compared, but they do different jobs. MCP is how your AI tool reaches the knowledge base, and RAG is how the knowledge base finds the right passages for it. In CoFix they come joined.
Explore Knowledge BaseRAG as a service
A RAG system, without building one.
Chunking and embeddings
The app chunks and embeds your pages itself, so there is no ingestion pipeline to write or keep running.
No vector database to run
Documents, the links between them and their embeddings are stored together as a graph, so there is no separate store to host.
A hosted MCP server
One MCP connector per member for Claude, ChatGPT, GitHub Copilot or Gemini CLI, and for other AI tools that support MCP.
Permission-aware retrieval
Each connector signs in as its member and searches only the knowledge bases they can read, with access set per team.
Only reviewed pages
New pages that fail the automated check are held for review and stay out of retrieval until someone approves them.
Answers you can trace
Every assistant action, in the app and over MCP, is in the activity log beside the edits people make.
Retrieval quality
Knowledge Health shows retrievals, ratings and match quality per document, and the searches that found nothing.
Freshness
Stale pages people still read are flagged, ranked by reads, with a button that asks the assistant to draft the update.
In depth
A closer look.
Graph RAG
Graph and vector search in the same query.
Documents and the links between them are stored as a graph, with embeddings for meaning. One search finds pages by what they mean and by how they connect, so a question about a deploy also reaches the runbooks it links to. Your AI tools get the same retrieval over MCP: graph RAG and vector RAG in one.
- Vector search by meaning
- Graph search by relationship
- Both in one query
- The same retrieval in the app and over MCP
Reviewed before retrieved
A draft that fails the check never reaches an answer.
Each knowledge base has a purpose and the sections every page must answer. A new page that leaves one empty, like a runbook with no rollback steps, is held in the review queue and kept out of search and retrieval until someone approves it. That matters more once AI agents write pages as well as read them.
- Required sections per knowledge base
- Pages that fail are held out of retrieval
- Hold every new page if you prefer
- Read, write or full access per team
RAG evaluation
See which documents your AI tools rely on.
Knowledge Health shows how each document does as a source: how often it is retrieved, how people rate the answers it gives and how well it matches the questions. It lists the searches that found nothing, so you know what to write next, and the stale pages still in demand, so you know what to fix first.
- Retrievals, ratings and match quality per document
- Searches that found nothing, ranked by demand
- Stale pages still in demand, ranked by reads
- A health score you can recompute whenever you like
Connects to
The assistants your team already uses.
Questions
RAG FAQ.
What is a RAG knowledge base?
A knowledge base that AI models retrieve from before they answer, so their answers come from your own documents rather than from what the model remembers. In CoFix every knowledge base is one: pages are chunked, embedded and linked by the app, and your AI tools search them over MCP.
What is a RAG MCP server?
An MCP server that does retrieval: an AI tool sends it a question over the Model Context Protocol and gets back the passages that answer it. CoFix runs one for your knowledge bases, with one connector per member, so there is nothing to host.
What is the difference between RAG and MCP?
RAG, retrieval augmented generation, means a model looks up your own content before it answers. MCP, the Model Context Protocol, is the open standard an AI tool uses to reach data and tools. Here they work together: MCP connects your AI tool to the knowledge base, and the knowledge base does the retrieval, so you get RAG without building it.
Does MCP replace RAG?
No. MCP is a way for an AI tool to reach data, and it doesn't decide which part of that data answers a question. An MCP server with no retrieval behind it leaves the AI tool to fetch whole documents and search them itself. RAG is the retrieval, and MCP is how your AI tools reach it.
Do we need a vector database?
No. Documents, the links between them and their embeddings are stored together in the knowledge base and searched with graph and vector search at once. There is no vector database, embedding pipeline or MCP server for you to run.
What is graph RAG?
Retrieval that follows the links between documents as well as their meaning. Vector search finds the passages that mean what the question means, and graph search finds the pages connected to them. CoFix does both in one query, in the app and over MCP.
Can Claude Code use our knowledge base for RAG?
Yes. Add your MCP connector to Claude Code and it searches your knowledge bases, tickets and work log as it works, signed in as you. It works the same way in Claude, ChatGPT, GitHub Copilot and Gemini CLI.
Can an AI tool see documents its user can't?
No. Each connector signs in as its member and searches with that member's access: read, write or full access per team and per knowledge base. Pages held for review stay out of retrieval until someone approves them.
Which AI tools can connect?
Claude, Claude Code, ChatGPT, GitHub Copilot and Gemini CLI, and other AI tools that support MCP, the open standard the connectors speak.
How is RAG over MCP priced?
It is part of Knowledge Base: a seat for each person who can reach Knowledge Base, plus storage per gigabyte. The pricing page has the numbers.
Part of one workspace