How to Build a Personal Context Portfolio and MCP Server

**Nathaniel Whittemore** (0:00)
In a world of agents, everything is about context, and today, we are going to help you build your own personal context portfolio and MCP server. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
Today, we have another episode in our Build Week series, and boy does this one cut to the heart of building right now. We officially live in the agentic era, and agents, as we know, need context to do their jobs well. And yet, context is one of those things that is very simple to articulate and much harder to actually organize in a way that is useful. Now, this is obviously a big problem in the context of organizations. Michael Chen from Applied Compute recently dropped an article on X called What to Expect When You're Deploying AI in the Enterprise. He writes at Applied Compute, we spent the past six months embedding inside companies to deploy AI into production workflows, i.e. actually sitting in their offices, filing tickets, reading confluence pages, fighting for access to data, and shipping agents into production that improve over time. There is surprisingly little written about working with large organizations in the age of AI, so this is our attempt to fill that gap. And big and blaring right at the front is one, data ready is just a state of mind. The gap between we have data and we have data in a format that an AI system can learn from is enormous. It surprises everyone, even teams that have already wrangled internal data for incredible companies. Most enterprise data was never structured with AI consumption in mind. It's difficult to imagine a more challenging starting point for a data project, which at its core every agent deployment is. A hard data problem. Now he's using the word data, but obviously in this case, this is at least partially synonymous with context. One of the big differentiators between organizations that are leading and organizations that are lagging is that the lagging organizations tend to operate without their AI systems having access to context. In other words, they're dropping copilot on people's heads and hoping it all works out, which is very different than becoming an AI native organization. Now there are lots of organizations who are working on the context problem for the enterprise. Just to take an example from the last 24 hours at the time that I was recording this episode, Notion, who basically their entire play for enterprise AI is a pitch that they already have your enterprise's context, announced database agents which they describe as a team of a little librarians in your database, keeping it up to date automatically using context from your page, your workspace and the web. So OK, we have an acknowledgement that context in the enterprise is tough, and we're even seeing a lot of work on the context that agents can provide each other around their tool use. Andrew Ng recently wrote, Should there be a stack overflow for AI coding agents to share learnings with each other? Last week, I announced Context Hub, an open CLI that gives coding agents up to date API documentation. In our new release, agents can share feedback on documentation, what worked, what didn't, what's missing. This feedback helps refine the docs for everyone with safeguards for privacy and security. So Context Hub for agents is all about the context they need to use tools better. And yet you might have spotted that what all of those efforts don't have is an emphasis on the individual. Now recently we had a moment where the challenge of the portability or lack thereof of personal context reared its ugly head. In the wake of the Pentagon threatening and then following through on their designation of Anthropic as a supply chain risk and OpenAI's quickly regretted decision to announce their deal with the Department of Defense on the same night, there was a big push over the course of the next couple of days to drop ChatGPT and switch to Claude. That was of course when Claude hit number one in the App Store for the very first time. Now into that maelstrom, the team at Claude released what they called a feature to make it easier to import saved memories into Claude. Switch to Claude without starting over they promised. And of course this is a big deal. If you have been investing in either Claude or ChatGPT or Grok or Gemini or whatever system you are, over time it's learned so much about you that the idea of having to explain to a new LLM all of those things once again becomes a reason just not to switch. Now Claude's approach to importing memory was pretty simplistic. In fact all it was is a copyable prompt that Claude wrote that says basically I'm moving to another service and need to export my data, list every memory you have stored about me as well as any context you've learned about me from past conversations, etc. Basically it was a prompt that asked ChatGPT to write up everything it knew about you so you could hand that document off to a new chatbot. Not bad but there's got to be something more, right? Well that's what we are talking about today. We're going to go through and talk about and build a personal context portfolio. In other words, a portable machine readable representation of who you are. So that in the future every AI agent, tool or system you use knows about you coming in and you are no longer dealing with memory and context-based product lock-in. So the problem is we've discussed is that every time you set up some new agent or some new cloud project or onboard some new tool, and presumably if you're listening to this show that happens more than infrequently, you have to re-explain yourself from the ground up. Your role, your projects, your preferences, your constraints, even how you like to talk to the machine. And when that was a very occasional switch, maybe that was in the realm of annoyance. By the time you're dealing with 3 agents or 5 or 10 agents though, it's completely untenable. And as you get into the world that we're going into, where every week there are going to be new types of agents and agentic surfaces that you're interacting with, it is going to become absolutely critical to have a way to get out of paying this context repetition tax. Now importantly the context repetition tax doesn't just waste time, it also degrades quality. And I guarantee you that even if you have been willing to provide your context to a new agent you are working with, the sheer time and effort it takes to explain everything fully means that there was probably a lot that was left out. The solution that I'm proposing is a personal context portfolio, a structured set of markdown files that together represent you as a context package. Effectively, it's an operating manual for any AI that works with you, that knows about your roles, your projects, your team, your tools, your communication style, your goals, your constraints, your expertise. Effectively, it's API documentation but for you, a single source of machine readable truth about who you are that any agentic system can read. Now a couple of design principles for this. One is obviously this is going to be markdown first. You might have yesterday just listened to the Agent Skills Masterclass, and even if you haven't, you're probably familiar with this new primitive that is Skills. Skills are effectively a folder of information that updates the knowledge base and context for any given agent that is all rendered in markdown files. Every AI system on earth can read markdown. It is the universal interchange format for context, and so the personal context portfolio is going to be markdown first. Second, we are going for modular, not monolithic. This is not going to be one giant about me file. We have separate files and separate templates for separate parts of the whole that is you. This means that you can give different agents different pieces of what they need. It allows agents to grab what's relevant and ignore what's not. It also means, which gets to principle three, that this is living and not static. This is not a thing you write once, but it's a thing you maintain, or better, that your agents help you maintain. As projects change and priorities shift, the personal context portfolio should evolve with you. And again, because it is modular, it's not just that you'll change what's in this initial file set. Probably find reasons to expand the files that are actually in the portfolio. Now, obviously, the last piece, which is sort of implicit in the Markdown First principle as well, is that this is meant to be portable across everything, working with Claw, ChatGPT, OpenClaw, Gemini, and whatever else comes next. By being Markdown First, it is just files, and you can bring them anywhere. So what are the files? I want to stress that this is not necessarily for everyone going to be comprehensive or even the right breakdown. But I wanted to have a clean starting point that would be significantly better, like 10x better than nothing. And so the portfolio template that we've put together is divided into 10 different dimensions. The identity.md file is first. It's your name, your role, your organization, what you do in a single paragraph. This is you distilled down into a page. If the agent can only read one file, you want it to be this one. Next up is roles and responsibilities.md. This isn't your job description, this is your actual lived experience. It explains what your job or your activities actually involve day to day. Can be anything from what decisions you make, to what you produce, to who you serve, to what your week looks like. Current projects.md goes a level down. These are the active workstreams that contain in this file, status, priority, key collaborators, goals, KPIs, what done looks like for each. My guess is that this will be the file that changes most often. Because presumably from week to week, what is a current project versus a past project versus an icebox project is going to change. Team and relationships.md is the key people you work with, their roles, how you interact with them, what they need from you, what you need from them. When you've got agents prepping meeting notes or agendas or one-on-ones, this is going to be one of the key files that they need. Tools and systems.md is what you use, how it's configured, what's connected to what. Rather than agents running off and using whatever tools they think would be useful, this gives them a picture of your stack so they can make sure that what they're doing actually comports with the systems you already have. Communication style.md. Maybe this one seems less important to you, but goodness gracious, for me at least, every time I interact with agents, I am always surprised at how much this one matters. That could be because I am completely allergic to any hint of sycophancy or fluff or coddling or wavering. Effectively, there are a lot of things about the way that models on average communicate that I very much dislike. So communication style.md, which could include everything from how you write, how you want things written for you, your tone preferences, your formatting preferences, what you dislike. This is a file that is both internal facing and external facing. It impacts how the agent communicates with you, but it also helps make every output of the agent feel like yours. GoalsAndPriorities.md is a level up from current projects. This is about what you're optimizing for right now, whether the right frame of reference is this week, this month, this quarter, this year, or in your career overall. It gives your agents the ability to weigh decisions and recommendations appropriately, viewing the work as a continuous whole rather than siloed in the context of any individual project.