Three terminal tabs open. Claude Code in one, reviewing a PR. Cursor in another, writing tests. A third session deploying to staging. Each one is fast, capable, and completely unaware that the other two exist.
So when the PR review finds a type error in auth.ts, the deploy session charges ahead anyway. The test writer generates assertions against a schema that changed twenty minutes ago. Nobody told them. Nobody could.
I was the coordination layer. Me. Copy-pasting context between windows like a short-order cook running plates between three kitchens that don't share a ticket system. The agents could write code in seconds, but they couldn't pass a note. Every "hold on, let me update the other session" was me admitting that the smartest tools in my stack were bottlenecked by the slowest component in the room.
That got old on a Tuesday afternoon in January. I'd just spent forty minutes manually synchronizing context across four Claude Code sessions that were supposed to be working on the same sprint. Forty minutes. I could've built a feature in that time. Instead I was a human message bus, and not a particularly reliable one.
So I built the message bus.
The name came from my living room. My two boys, Bastian and Caspian, were tearing around the house while I was heads-down on the first prototype. They asked what I was building. I looked at them and said, "How about CacheBash?" Cash and Bash. Their nicknames.
Then it hit me that the name actually works. We send bash commands to the cache. We bash the cached bash. Sometimes the best names find you while your kids are climbing on your back.
CacheBash
An MCP server. You point your AI client at it, hand it an API key, and your sessions can coordinate without you standing in the middle.
The architecture is what you'd expect if you've built distributed systems: Cloud Run server in TypeScript, Firestore for persistence, MCP protocol with Bearer auth for client connections, React Native mobile app for when you need a human in the loop but don't want to be chained to a terminal.
What matters is what it lets agents do.
Dispatch is the task layer. One session creates a task with a priority. Another session claims it, works it, marks it done. Claim-based ownership means two agents never grab the same work. Priority queues with interrupt, queue, and backlog actions mean urgent things get handled before routine ones. The basic primitive that every multi-agent system needs and almost none of them have.
Relay is the communication layer. Direct messaging between sessions. Multicast to groups. Threaded conversations with delivery tracking. That code review agent that found the type error? It sends the finding directly to the test writer and the deploy agent. The deploy agent pauses. The test writer adjusts. Nobody relayed anything except the agents themselves.
Sprint orchestration is where it compounds. You define stories with dependencies. "Run integration tests" depends on "fix type errors" which depends on "complete PR review." The server schedules them in waves, runs parallel work within each wave, and advances to the next wave when the current one clears. You watch progress from your phone while doing something else entirely.
There are 34 tools across 9 modules. Program state for persistent memory between sessions. A Signal module that pushes decisions to your phone when an agent needs human input (approve a deploy from the grocery store, pick a migration strategy from the couch). Dream scheduling for autonomous overnight work with budget caps. Full observability with audit logs, cost tracking, and execution traces.
I'm not going to walk through all 34. The docs do that. The point is: the agents coordinate through the server, not through you.
What It Looks Like
Add this to your MCP config:
{
"mcpServers": {
"cachebash": {
"url": "https://your-instance.run.app/v1/mcp",
"headers": {
"Authorization": "Bearer YOUR_API_KEY"
}
}
}
}
Restart your client. Create a task in one session:
> create a task titled "Review auth module" with priority high
Claim it in another:
> check for pending tasks and claim the highest priority one
Two sessions. Coordinated. No shared filesystem, no custom scripts, no clipboard.
Scale that up. A review agent finds three issues and creates a task for each, tagged with priority and file path. A builder agent polls for tasks, claims one, fixes it, marks it complete. A third agent watches for all tasks to resolve, then opens the PR. Chain the whole thing through sprint orchestration and you've got a team of agents running a real development workflow while you're on a walk with your kids.
Program state makes it durable. Each agent writes what it was working on, what it learned, and what's still unfinished. When a session ends and a new one starts hours later, it reads that state and picks up mid-task. The context lives in Firestore, not in a chat window that evaporated when you closed the tab.
Compatibility
CacheBash speaks Streamable HTTP transport with Bearer token auth. Any MCP client that supports both works out of the box.
Tested: Claude Code (daily driver), Cursor, VS Code with Copilot, Gemini CLI. Four out of five major clients, no backend changes needed. ChatGPT Desktop is the holdout; it requires OAuth 2.1, which is on the roadmap.
What's Coming
OAuth 2.1 to unlock that fifth client. Webhook integrations for piping task completions into Slack or Discord. GitHub sync so tasks and sprints mirror to Issues and Milestones automatically.
The MCP ecosystem is growing fast. Every week, another client ships MCP support. CacheBash doesn't care which one you pick.
The Repo
Open source under MIT. Self-hostable on GCP. The server, mobile app, and all 34 tools are in the repository. Deploy to your own Cloud Run instance and your data stays yours. No third-party accounts. No SaaS dependency.
The mobile app is in TestFlight now (iOS). Request access from the repo README. Android and App Store release are coming.
If you're running multiple AI sessions and coordinating them by hand, you're doing it wrong. The agents are capable. The protocol exists. The infrastructure is free and open. The only thing missing was a way for them to talk to each other, and that's not missing anymore.
Send me a message and let's talk.