Nowadays, in the world of generative and agentic AI, where we as humans-in-the-loop need to process tons of information and content, and need to produce and control even more data with our AI-assisted tools and systems (internally and externally) there is no doubt that data and knowledge are a very important asset, or even a resource.
The quality of that knowledge defines the relevant context in which agents operate. It could be inside your own loops and tools, or inside the wired-up Claude or ChatGPT (or whatever we use). The context defines the results and output of our agents. And the output of our agents defines the quality of what we bring out externally: content, conclusions, programming code, etc. So knowledge is the cornerstone of every business, digital solution, service and expertise.
It's a new layer that sits close to the relational data we keep in our databases. But it is not the same. It's important to differentiate between the data produced by a business system (whatever business processes it operates) and the knowledge data that explains how and why those business processes operate. Relational data answers the questions WHAT and WHO, while knowledge data answers the questions WHY and HOW. WHAT and WHO are the classical records of data in every system, but WHY and HOW are something new for systems. They were always a hidden asset located in endless Google Drive and Confluence documents, in onboarding instructions, separate specifications and pitch decks. It was always real knowledge, settled in real people's brains.
Now there is a necessity to bring this knowledge into the system, because the answers to WHY and HOW are now needed to operate AI-assisted processes: from planning a new feature in code, to analyzing cash flow reports, to building next month's content plan sliced by particular ICPs for specific channels.
So there are clearly two related tasks to solve: organize the knowledge and connect the knowledge to AI.
And what is interesting, how you organize the knowledge influences how you connect it to AI.
A crowded space with hard questions
Of course, this is already a solution-crowded space. But there are some conditions that impacts the solution: how do you want to control the knowledge as a single source of truth, and distribute it, wired up to AI tools, equally and in a controllable way to different humans in the loop?
The easiest setup is just to create a project in a Claude Team organization, for example, and upload the necessary documents or wire up Google Drive. But how do you manage changes? Who controls which document should be added next? How do you sharpen context enrichment out of this endless sea of textual content? And what if we need this not in Claude, but inside a CRM or ERP system, where an AI assistant should answer questions related to that system's data?
If we stay in Claude or ChatGPT, how do we distribute it equally to design, code and other chats?
- How do we control what data we've added to our AI assistant?
- How do we make revisions and version it?
- How do we immediately revoke access to that data for certain connectors or users?
- How do we observe which agent asked about the knowledge, when, and with which human in the loop?
- How do we make search inside that knowledge good?
- How do we handle all of this? Maybe just put everything into a vector database? And then how do we manage those?
- What if we need this internally today, and tomorrow need to apply the same approach inside a particular system, or adapt it to a custom client setup?
Knowledge as a codebase
So, answering all of that: knowledge that can be counted as an organizational or business asset should be kept and managed like a codebase. As example of this asset, it could be project specifications, user guidelines, onboarding guides, API documentation, marketing strategy, a specific glossary, brand identity, ERP/CRM functionality description, etc. A vector database can still be useful for search, but only as a derived index built on top of the knowledge which you can rebuild at any time not as the place where the knowledge itself lives.
And this is where Git is the answer. It already has all the necessary attributes for this case: versioning, a single source of truth, and a merge/pull process. Each knowledge asset can be organized as a separate Git repository with the necessary folder structure and files in Markdown (.md) format, with attributes, wiki links between documents, etc. — the perfect material for AI to read, understand and keep as context.
Connecting it to AI through MCP
In this setup, we can have a tiny, secure, self-hosted web application with an MCP server exposing search and read tools to the agents.
And this web app is the whole point. Most knowledge owners like marketing, brand, sales, operations don't use Git, and they shouldn't have to. The web app gives them a simple editor on top of the same repositories: they browse, edit and add Markdown documents in the browser, and every change is committed to Git behind the scenes, with the same history and revert as any developer commit. Technical people can still work with the repository directly from their IDE or Obsidian. Both roads lead to the same single source of truth.
We can connect to the MCP server via OAuth or a token, depending on the external AI tool. Claude, for example, works perfectly with OAuth, where the connection is bound to an exact human in the loop. In other cases, a token-based MCP connection is enough. For custom connections, we can also restrict access to specific IP addresses (for example, if we know the connector will only be used from a particular production ERP server).
Each MCP connector maps to one knowledge repository and is bound to its Git repository with an SSH key. When an agent connects to a knowledge repository, the web app (if it hasn't already fetched and cached it recently) git-clones the Markdown files from the configured branch of the configured repository using the SSH key and brings them into storage. They are then available for the agent to read and search. Every agent action is recorded: which tool was used, by which human in the loop, and when.
The result
And then it works perfectly: every human using an AI-assisted tool always has access to the freshest git-pulled version of a particular knowledge asset. For the organization it's also beneficial: each knowledge asset is set up, controlled and observed, and all connections from all tools are logged.
At Well Digit we built such a solution for our internal setup and are very happy with the results.
Facing the same questions? If you're thinking about how to organize your company's knowledge, keep it under control as a single source of truth, and connect it securely to Claude, ChatGPT or your own CRM/ERP, I'd be happy to talk. We can consult on the approach or help you build and set up a knowledge-as-code infrastructure for your team or product.
Feel free to reach out to us on our Contact page, and check out our self-hosted solution Notedigit, where we solve this task in exactly the way described above.