Hey HN, we're Jonathan, Ollie, and Lucas, and we're building Compendium ([https://cerenovus.ai/compendium](https://cerenovus.ai/compendium)). Compendium is a shared workspace that keeps your whole team, humans and coding agents alike, working from the same source of truth.
We started building this after running into the same wall over and over: AI made shipping features fast, but it didn't make coordinating them any faster. One of us would build a feature only to find a teammate had built an identical one. We'd make an architectural decision while someone else was making a conflicting one at the same time. And every agent session started from zero, so we were pasting the same context about our codebase and our decisions into Claude Code dozens of times a day. If you 10x the number of features you ship without 10x'ing the number of meetings you have, things drift out of sync pretty quickly. Add agents that each need the full picture, and it gets out of hand fast.
Here's how it works. Every document in Compendium is markdown, but under the hood each one is a node in a permissioned knowledge graph, and links between documents are edges. The whole graph is exposed over an MCP server, so any agent you point at it (Claude Code, Codex, or anything else that speaks MCP) can traverse your team's decisions, ownership, and the rationale behind them. That last part matters to us: agents shouldn't just retrieve documents, they should be able to answer "why did we choose this?" and not just "what did we choose?"
Real-time collaboration on documents runs on Yjs CRDTs over CodeMirror 6, yielding a smooth context editor that feels like if Obsidian and google docs had a baby. It's perfect for going over specs, especially given that you can run our builtin AI agents with all your context to markup a spec and suggest revisions. But we also took this idea of real-time collaboration and used it to make our favorite feature: multiplayer agent sessions. Two people can jump into the same session and drive one Claude (or Codex) together, like a Google Doc for AI, letting you quickly hand-off work or collaborate on coding. There's also a live (opt-in) view of what every teammate and agent is working on right now, plus summaries of what changed while you were away. Stack-wise it's Next.js on the frontend, Node/Express and Postgres on the backend, and a zero data retention posture where the model provider is treated as pure compute: all state lives in infrastructure we control.
*Last note - one feature we just pushed is that I think is really cool is that the whole app is agent-drivable, meaning almost every action you could take in the app is also possible via the MCP server (when authorized)- including driving your Claude sessions and pulling up files for you to look at.
We've been dogfooding it for all of our internal coding for a few weeks now, and the honest report is that the number of "oh, I thought you were doing that" moments has dropped to roughly zero, and our agents stopped asking us questions the wiki already answers. It still feels kinda magical to see it in action.
You can try it at [https://cerenovus.ai/compendium](https://cerenovus.ai/compendium). There's a 2-week free trial, no sales call, no onboarding meeting, you can just sign up and point your agents at it. We'd love your feedback, especially the critical kind, and we'll be in the thread all day answering questions.
Hey HN, we're Jonathan, Ollie, and Lucas, and we're building Compendium ([https://cerenovus.ai/compendium](https://cerenovus.ai/compendium)). Compendium is a shared workspace that keeps your whole team, humans and coding agents alike, working from the same source of truth.
Quick launch vid here: [https://www.youtube.com/shorts/Vw\_aFL7YOUI](https://www.youtube.com/shorts/Vw_aFL7YOUI)
We started building this after running into the same wall over and over: AI made shipping features fast, but it didn't make coordinating them any faster. One of us would build a feature only to find a teammate had built an identical one. We'd make an architectural decision while someone else was making a conflicting one at the same time. And every agent session started from zero, so we were pasting the same context about our codebase and our decisions into Claude Code dozens of times a day. If you 10x the number of features you ship without 10x'ing the number of meetings you have, things drift out of sync pretty quickly. Add agents that each need the full picture, and it gets out of hand fast.
Here's how it works. Every document in Compendium is markdown, but under the hood each one is a node in a permissioned knowledge graph, and links between documents are edges. The whole graph is exposed over an MCP server, so any agent you point at it (Claude Code, Codex, or anything else that speaks MCP) can traverse your team's decisions, ownership, and the rationale behind them. That last part matters to us: agents shouldn't just retrieve documents, they should be able to answer "why did we choose this?" and not just "what did we choose?"
Real-time collaboration on documents runs on Yjs CRDTs over CodeMirror 6, yielding a smooth context editor that feels like if Obsidian and google docs had a baby. It's perfect for going over specs, especially given that you can run our builtin AI agents with all your context to markup a spec and suggest revisions. But we also took this idea of real-time collaboration and used it to make our favorite feature: multiplayer agent sessions. Two people can jump into the same session and drive one Claude (or Codex) together, like a Google Doc for AI, letting you quickly hand-off work or collaborate on coding. There's also a live (opt-in) view of what every teammate and agent is working on right now, plus summaries of what changed while you were away. Stack-wise it's Next.js on the frontend, Node/Express and Postgres on the backend, and a zero data retention posture where the model provider is treated as pure compute: all state lives in infrastructure we control.
*Last note - one feature we just pushed is that I think is really cool is that the whole app is agent-drivable, meaning almost every action you could take in the app is also possible via the MCP server (when authorized)- including driving your Claude sessions and pulling up files for you to look at.
We've been dogfooding it for all of our internal coding for a few weeks now, and the honest report is that the number of "oh, I thought you were doing that" moments has dropped to roughly zero, and our agents stopped asking us questions the wiki already answers. It still feels kinda magical to see it in action.
You can try it at [https://cerenovus.ai/compendium](https://cerenovus.ai/compendium). There's a 2-week free trial, no sales call, no onboarding meeting, you can just sign up and point your agents at it. We'd love your feedback, especially the critical kind, and we'll be in the thread all day answering questions.