Let's consider a five-person firm. Maybe a consulting firm (as was the build by @Eytan Levy ). Each consultant has a workspace of markdown files, client context, working notes, output logs, and Claude Code runs on top of it. The whole thing lives in one company repo. Everyone clones it, pulls in the morning, pushes at the end of the day. It works. I'd build it the same way. Now ask what's stopping consultant three from opening client seven's folder. The honest answer is nothing. She just doesn't. There's a contract, there's training, she's a professional, and you know her. That's a real control and I'm not going to pretend otherwise. But it's worth being precise about what's actually holding the system together, because it isn't the architecture. Git permissions stop at the repo boundary. Inside the repo, everything is readable by everyone who cloned it. The thing keeping client seven private is not a permission. It's a person. That costs you nothing while it's true. The problem is that nothing in the system tells you the day it stops being true. And it stops for ordinary reasons. You hire past the point where you personally know everyone. You bring a contractor in for one engagement. Someone leaves for a competitor. You sign a client whose contract says their data can't sit on personal devices. You take on a client who competes with a client you already have, and the gap between two folders suddenly needs to be an actual wall. On the day one of those lands, you reach for the lever and find out what you really have. You remove the person from the repo. That stops their next pull. The clone they made three weeks ago is on their laptop, complete, and it stays there. Repo permissions govern who can get it. They don't govern who has it. That is distribution control, and most people think they bought access control. The part I keep my eye on is calling this a big-company problem. It isn't a headcount thing. A fifty-person firm of salaried employees under proper contracts might be perfectly fine. A three-person shop with one contractor who also works for your competitor is not. The trigger isn't scale. It's the first time there's someone inside the system you wouldn't personally vouch for. That usually shows up with growth. It doesn't have to.

I built an internal AI workspace for a consulting firm (5 users). Each consultant has a workspace with structured markdown files (client context, skills, stages, output logs) — all orchestrated by Claude Code locally. The filesystem IS the database — folders = tables, .md files = records. It works perfectly on local machines. Now I need to take it online so the company owns the data centrally (not on individual PCs). Questions: 1. Is a filesystem-based architecture viable at this scale or do I need a real DB? 2. Can I connect a web interface directly to a centralized filesystem? ( Should i use VstudioCode Web or Recreate an Chat interface ) 3. How do you handle access control with file-based data? 4. What's the best way for users to interact with the agent + data remotely — VS Code web, custom UI, something else? Objectif are : The Company Own the Data - Agent Claude Code

where can I ask my questions for the afternoon tea ?

Recording's up, and I packaged the whole thing so you can drop it straight into your second brain or hand it to your AI. 🍵 Here's what we got into this session: 🧠 Map your work, don't just store it. A second brain holds notes. A map holds your work plus the people and data around it — teams, processes, and the links between them. You can't improve what you can't see, and neither can your AI. 🔗 Every workflow is a node. One markdown file = one process. The references between them are the edges. That's the whole graph. 🟢 Build workflows, not outputs. Store the outputs inside the workflow. Then when Opus 4.8 or Fable ships, the right feeling is "cool, my system just got better" — not scrambling. 🔍 Google basically proved the method. Their new Open Knowledge Framework (dropped June 12) is markdown + files + front matter to document and query big datasets. One of the biggest players outside Anthropic is doing the markdown-and-files approach we've been practicing here. 👀 🏛️ Plus: mapping a real company's teams in Obsidian, the platform coming so you can own and license your workflows, and a sneak peek at my new paper — Human in the Compute Layer — built on Engelbart's 1962 work. 📎 What's attached (and what each file is for): 📝 session-notes.md — The opinionated version. All the ideas from the call, written so you can act on them. Start here if you want the short version. 📚 term-sheet.md — Plain-English definitions for every term: node, edge, semantic layer, OKF, ICM, "the data becomes the agent," and more. Perfect if you're new to the room. 📄 vault-page.md — The index for the whole package. 🗂️ Package.zip — Everything zipped, ready to add to your AI's memory (Claude, Hermes, OpenAI — whatever you run). 💬 Watch it, grab the files, and drop your questions below — the best ones seed the next Afternoon Tea. So much love. 🫶

Do someone have tried claude Tag. I would like to know the difference between Claude Tag and the ICM methodology. Maybe antrhopic with Claude Tag has find a scalable solution for a clear context, and memory that evolve with time ? By the way, how did you do to make an memory evolve with new information? For exemple : T0 : your knowledge.md contain a specific information about something so this information is true 👍 T1 : after some research your knowledge.md evolve but the specific information that you find in not correct T2 : how does your system know which information is true and accurate ? Do you understand what i mean ?