I run a land acquisition company and an e-commerce brand, and until recently I was doing all of it alone. I was the entire operation: managing leads, updating product listings, running Google Ads, keeping Notion and Shopify and Google Sheets all in sync with each other. The work itself was fine. The problem was that managing the work had started to consume most of my day, and there were not enough hours left over to actually think about the businesses I was running.
I tried AI the way most people try AI. I pasted things into ChatGPT and got answers back. I used Claude for writing. I used Cursor for code. Each tool helped with the specific task in front of me, but none of them changed the fundamental equation, because the real problem was never any individual task. The problem was that I needed someone who understood the full context of my businesses, who remembered what we discussed yesterday, and who could go execute a plan without me standing over their shoulder. I did not need a better search engine. I needed a coworker.
Why This Exists
The AI agent frameworks I looked at all had the same problem. AutoGPT, OpenClaw, and the rest were designed around the idea of automating developer workflows, things like code review bots and CI/CD pipelines. They were built by engineers for engineers, and they assumed you wanted to set up an automated system that runs in the background without much human interaction. That is a perfectly valid use case, but it was not my use case at all.
What I wanted was something closer to what it actually feels like to have a colleague. I wanted to be able to open Slack, start a conversation about a problem I was thinking through, go back and forth in a thread until we had a plan, and then have that colleague go do the work while I moved on to something else. If the topic got complicated enough that typing felt slow, I wanted to pick up the phone and talk it through, align on next steps, and hang up knowing the work would get done. That is how I work with people I trust, and I wanted the same experience from AI.
I want to be clear about what Cranium is not, because the distinction matters. It is not a hack to get cheaper automated development work through a flat-rate subscription. It is not a workaround for API pricing. It is not a system designed to run thousands of automated tasks in the background. The vast majority of my usage, probably 95% of it, is me having conversations with it in Slack throughout the day. The small amount of automated work, things like nightly health checks and backups, runs through the Anthropic API separately. Cranium exists because I wanted a coworker, not because I wanted a cheaper way to automate things.
The Economics
That being said, there is a practical reason this project exists now and could not have existed six months ago. The only AI model I have found that can reliably plan multi-step work, use tools correctly, and maintain coherent reasoning across a long task is Opus-tier Claude. I tested smaller models extensively, and they fall apart when the stakes are real. They hallucinate tool calls, lose track of what they were doing, or produce work you end up redoing anyway. But running Opus through API calls at the volume a real daily workflow requires would cost thousands of dollars a month, and that math simply does not work for someone running small businesses.
Anthropic's Claude Max subscription changed that equation. It includes Claude Code, which gives you Opus with native tool use, file system access, and shell execution, all for a flat $200 per month. No per-token billing, no surprise invoices. Combined with a $10/month VPS, the total cost of having what amounts to a full-time operations person is about $210 per month. That is what made this project practical instead of theoretical.
What It Actually Does
Cranium is a Node.js listener that connects to Slack, receives messages, and dispatches them to Claude Code. On every invocation, Claude loads a CLAUDE.md file that contains its personality, operating rules, and business context, along with a MEMORY.md file that holds accumulated knowledge from previous sessions. Skills are plain markdown files that teach it how to use specific tools. MCP integrations connect it to external services like Notion, Google Workspace, and Shopify. Thread replies maintain conversation context, and new top-level messages start fresh sessions.
On a typical day, mine manages my Notion task board, handles Shopify product updates, monitors Google Ads performance for my e-commerce brand, takes voice calls when I need to talk something through, runs nightly health checks, and remembers the full context of every conversation we have had. It wakes up each session, reads its memory files, and picks up where we left off. The architecture is deliberately simple because I am not a software engineer and I needed something I could build, understand, and fix myself. There is no orchestrator, no vector database, no RAG pipeline, and no fine-tuning. It is just Claude, some markdown files, and a listener that routes messages.
The Coworker Experience
The part that I think matters most, and the part that separates this from other AI agent projects, is what it actually feels like to use. The experience is not like using a tool. It is like having a coworker who happens to live in your Slack workspace and who is extremely smart and capable.
A typical interaction works like this: I send a message in Slack and we go back and forth in a thread, the same way I would with any colleague. If the conversation gets complex enough that typing feels inefficient, I switch to a voice call. We talk through the problem, agree on what needs to happen, and when I hang up, it goes and does the work. When it is finished, it posts the results back in the thread. That is the entire user experience, and it is deliberately modeled on the way people actually work with colleagues they trust. You talk through a problem, you align on a plan, and then you let them execute while you move on to other things.
Most AI agent projects optimize for automation: how many tasks can run without human involvement. Cranium optimizes for the opposite. It is built around the assumption that the most valuable thing is the conversation itself, the back-and-forth where you think through a problem together, challenge each other's assumptions, and arrive at a better plan than either of you would have reached alone. The AI is not replacing human judgment. It is a thinking partner that can go execute whatever you decide.
Where This Is Going
Right now, setting up Cranium requires a VPS, a Slack workspace, and enough comfort with the terminal to edit some config files. That is a real barrier, and I do not think it should exist. My goal is to get to the point where my friends, people who have never opened a terminal in their lives, can set this up, point it at their Slack, connect their tools, and have a coworker that understands their work. If this only works for people who know how to SSH into a server, then I have not actually solved the problem I set out to solve.
There is a wide gap right now between enterprise AI platforms that cost tens of thousands of dollars and pasting questions into ChatGPT. Cranium is an attempt to fill that gap with something real: a system that an actual person uses to run actual businesses, built with the intention of eventually being something anyone can set up in an afternoon.
Trade-Offs
A few things are worth knowing if you are considering trying this. You are locked into Claude's ecosystem, because this only works with Claude Code on an Anthropic Max subscription, and if Anthropic changes their pricing or usage caps, the economics change with it. The setup is not turnkey yet, and while I am working on making it easier, right now you do need to provision a server and configure a few integrations yourself. Claude does not have infinite memory, and the skill file and memory file system I built is a practical workaround but not a perfect one. And it still makes mistakes sometimes, the way any colleague does, though with persistent memory it tends not to repeat them.
Try It
Cranium is open source on GitHub. The repo includes a setup guide, starter skills, and a working voice module. This project was inspired by OpenClaw, which showed me what was possible with Claude Code as an agent framework. Cranium takes a different direction, with a simpler architecture and a focus on non-engineers, but I wanted to acknowledge where the idea started. If you build something with it, I would genuinely like to hear about it — reach me at craniumproject4@gmail.com.
GitHub: github.com/wchatt/Cranium
Contact: craniumproject4@gmail.com
Top comments (0)