
caseOS
A knowledge graph that gives a case team’s AI agents the context of the case.
- Role
- Created and architected it; led the build
- Started
- As a side project on my laptop
- Team
- Grew to 10
- Launched
- Firmwide at BCG, May 2026
- Adoption
- Thousands of colleagues every week
- Works with
- Claude, ChatGPT, and other agents
Claude knows everything about the world and nothing about my case.
Ask it to help prep tomorrow’s meeting with Joe, and it stalls. Who is Joe? What did we cover last time? What does he care about? Did he send something ahead? A consultant knows all of this from being in the room. The agent knows none of it, and nobody has time to paste it in.
That gap grows as teams do more of their work with AI. A case produces a flood of context: transcripts, notes, email, client documents, steering-committee decks. The part that matters for any one task is small next to the whole, but still far too big to paste.
The idea
Give the agent a memory shaped like the work.
caseOS is a knowledge graph built around how a BCG case actually runs. It holds the clients, workstreams, hypotheses, meetings, and risks, and the relationships that link them, so an agent can come at the case from any angle and pull only what its task needs. As the team works, its agents read from that memory and write back to it.
Three decisions that shaped it
Bring the context to the agent people already use. Most AI tools ask you to come to their app. caseOS works the other way. It is built API-first and MCP-first, so Claude, ChatGPT, or a team’s own agent can use it directly. The website exists so people can see what their agents see.
Let agents write as well as read. Drop a transcript into a working session, and the agent updates the case. New people, decisions, and risks land in the right place, linked to what was already known. Every change reaches the whole team at once, so people and agents work from one real-time picture of the case. A shared memory that agents keep current was new when we built it, and it is still rare.
Respect who knows what. Knowledge at a firm like BCG lives at several levels: a person, a case, an account, a sector, a practice, and the whole firm. caseOS uses one ontology across all of them and keeps the levels distinct. Because the ontology is solid, the information inside it can stay messy. A team can add private case detail to a client the firm already knows, then promote what is worth keeping when the case ends. At the firm level, caseOS holds the abstracted version of a framework or methodology, above any single applied instance, so it can be maintained incrementally, wiki-style. The next team starts ahead.
How an agent finds its way
A new agent starts with an overview of the case’s main themes. From there it searches, opens what looks relevant, and follows the connections, much as a person would. Recency and importance help it decide what to read first. It stops when it has enough. The loop draws on two ideas from cognitive psychology, spreading activation and information foraging. Weighted connections add centrality-based search, so agents can find the most connected knowledge across a large body of material.
People get the same experience. Every entity reads like a Wikipedia article, and every reference is a link. As a kid I loved the Wikipedia game, getting from one article to another only by clicking links. That is how I wanted a case to feel.
From side project to the firm
I built the first version of caseOS on my laptop and kept going through a stretch with no official sponsor. Once the firm backed it, I led its rebuild for production, and the team around it grew to 10. I owned the architecture and wrote much of the code. The production version runs on AWS, using Neptune and OpenSearch. I built the engine its agents run on, using the Claude Agent SDK.
A tool that holds case knowledge has to earn the firm’s trust. We took caseOS through security, legal, privacy, compliance, and responsible-AI review, plus an external penetration test. Permission-aware access, isolation between teams, and defenses against prompt injection are built into the product.
caseOS launched firmwide in May 2026, eight months after the production rebuild began, and reached thousands of users within seven weeks. It has kept growing since, and today thousands of colleagues use it every week. Case-y, the AI teammate I led next, can draw on it too.