Hi everyone! I've been building this for a while and would love to hear what you think.
Cookbook started as a shared workspace (with shared memory) for a team and their AI agents, think Obsidian plus Google Drive, optimized for AI. It started off as a simple MCP server you added as a custom connector to Claude or Codex.
Original features included:
- Upload the files you're working on and they're extracted to markdown so agents read them efficiently.
- A shared memory attributed per teammate and per agent.
- Revision history so you can see whose agent did what.
Then I noticed, the model was no longer the important part. Models are constantly changing. They should be used as the “engine”; what matters is everything around it: files, memory, who did what and what it cost, etc. At the same time, being a student on campus I realized that there was no shortage of ambition around me. Students wanted to build (with friends), but many of them didn’t know where to start. What subscription do I get, what's the cheapest option, how can I work with my friends, are all questions I was hearing.
So Cookbook became a harness for that, bringing all your subscriptions, and teammates to one place. I designed the “Bridge”: a small daemon that runs your Claude Code or Codex on your own machine under your own login github.com/dpro10/cookbook-bridge. The chat was moved to Cookbook, along with the “room” where you can draw how your agents work together and watch them do it. Every run is designed to come back with a receipt showing what your agent did, and what it cost.
As students have started using it, and I have continued to build, here are some uses that have caught people's attention:
- “Subscription pooling” allows users to prompt all of their subscriptions in one place. If my teammate runs out of credits on his subscription he can switch mid thread and continue where he left off on mine (with my approval of course).
- Switching models has never been easier, simply connect another to Cookbook and it has all of your files, memory, and history.
- A robotics team connected Cookbook to NVIDIA Isaac Sim. Their agents can hand tasks to robots in simulated environments.
Some rough edges and things I’m still working on:
- The desktop apps are beta, the web app is how it should be used for now.
- Agents visiting another teammate’s hardware is still in the works
- Teams being able to train their own open weight models with their Cookbook runs is something I would like to work towards as well.
If you have any questions feel free to reach out, or point your agent to the “Start Here” workspace in Cookbook.
Hi everyone! I've been building this for a while and would love to hear what you think.
Cookbook started as a shared workspace (with shared memory) for a team and their AI agents, think Obsidian plus Google Drive, optimized for AI. It started off as a simple MCP server you added as a custom connector to Claude or Codex.
Original features included: - Upload the files you're working on and they're extracted to markdown so agents read them efficiently. - A shared memory attributed per teammate and per agent. - Revision history so you can see whose agent did what.
Then I noticed, the model was no longer the important part. Models are constantly changing. They should be used as the “engine”; what matters is everything around it: files, memory, who did what and what it cost, etc. At the same time, being a student on campus I realized that there was no shortage of ambition around me. Students wanted to build (with friends), but many of them didn’t know where to start. What subscription do I get, what's the cheapest option, how can I work with my friends, are all questions I was hearing.
So Cookbook became a harness for that, bringing all your subscriptions, and teammates to one place. I designed the “Bridge”: a small daemon that runs your Claude Code or Codex on your own machine under your own login github.com/dpro10/cookbook-bridge. The chat was moved to Cookbook, along with the “room” where you can draw how your agents work together and watch them do it. Every run is designed to come back with a receipt showing what your agent did, and what it cost.
As students have started using it, and I have continued to build, here are some uses that have caught people's attention:
- “Subscription pooling” allows users to prompt all of their subscriptions in one place. If my teammate runs out of credits on his subscription he can switch mid thread and continue where he left off on mine (with my approval of course). - Switching models has never been easier, simply connect another to Cookbook and it has all of your files, memory, and history. - A robotics team connected Cookbook to NVIDIA Isaac Sim. Their agents can hand tasks to robots in simulated environments.
Some rough edges and things I’m still working on:
- The desktop apps are beta, the web app is how it should be used for now. - Agents visiting another teammate’s hardware is still in the works - Teams being able to train their own open weight models with their Cookbook runs is something I would like to work towards as well.
If you have any questions feel free to reach out, or point your agent to the “Start Here” workspace in Cookbook.