The codebase is the thing you cannot reach#
Away from a laptop, you can read a ticket, reply to a thread, look at a dashboard. The one thing out of reach is the code, which is the thing every real question is actually about. Does this service still write to the old table? How much of the billing logic would this touch? Did we ever finish the migration?
So the answers get postponed until you are back at a desk, and the thinking that could have happened on a walk happens instead at the moment you least want to spend on it.
Rock8Cloud closes that gap. Connect it to Claude over MCP, and the connection lives on your Claude account rather than on a machine, so Claude on your phone can reach the same repository your team works in.
Ask, and something reads the actual repo#
Your phone does not grep your repository. It dispatches an agent that does.
Ask a question and task_agent starts a research session: a read-only agent with a sandboxed clone of the repo, which reads the real code and reports back. Not a summary of what you told it, not a guess from a model’s memory of similar projects. The files as they are on that branch today.
Runs take minutes and come back asynchronously, and on a phone that is the right shape. You ask on the way to the tram and read the answer three stops later. The question does not block you and neither does the answer.
ViewUse agent to ask about the stack in the web application service.
Task Agent
Alright, that one's queued up too. So now you've got two agents working in parallel, one on the API libraries, one on the web stack. Want me to poll both of them now and see how far along they are?
ViewTask Agent
Done, the agent's off and running, digging through the API code. It's queued right now, so it'll take a little bit. Want me to check back on it in a moment, or is there something else you'd like to look at while it works?
Use agent to ask about the stack in the web application service.
Two research agents are now reading two different services at once, each in its own sandbox.
ViewYes. Check the progress.
Get Agent Run · 2 steps
The API agent finished. The web framework is Elysia, with a CORS plugin alongside it. For the database it uses Drizzle as the ORM with the postgres driver, and Drizzle Kit for migrations. Authentication is handled by better-auth. On the tooling side it is all TypeScript running on Bun, with oxlint for linting.
The web stack agent is still running, though. Want me to check that one again in a moment?
ViewGet Agent Run
Your web app is a TanStack Start application, the full-stack React meta-framework, running React 19 with TanStack Router for file-based routing and TanStack Query for data fetching. There is no separate state library, TanStack Query covers it.
For styling it is Tailwind 4 with shadcn UI components built on Base UI, plus Lucide icons. It talks to your API through a type-safe Eden client, and auth is better-auth again. Like the API, it is built with Vite and runs on Bun.
Those four screens are one conversation, held entirely on a phone. The prompt goes out by voice, two research agents read the repository in parallel, and the answer that comes back names the actual libraries on the actual branch.
Brainstorming stops being reconstruction#
Planning away from a screen usually means planning from memory. Half the conversation is spent reconstructing what the system does, and the other half is spent hedging because nobody is sure.
With the codebase in the conversation, the disagreements resolve themselves. Claude can send a research agent to check how something is wired, pull the code review findings on the open branch, or read the runtime logs from last night’s incident. You are arguing about what to do, not about what is true.
That is the difference between a chat that produces opinions and a chat that produces a decision.
The decision becomes a pull request#
When the plan is good, the same conversation starts the work. task_agent again, this time as a coder session: its own sandbox, its own checkout, and it pushes a branch and opens a pull request with a preview environment.
Write the task as if the agent knows nothing about your conversation, because it does not - each session sees only its own prompt. Then carry on with your day. The work is waiting as a PR when you next sit down.
The gap between deciding something and it being underway collapses to one message.
What else is in reach#
Because the whole platform is on the same connector, the supporting questions are answerable too.
What is running. Projects, services, environments, deployment status, uptime and response times.
Why it broke. Build logs when a pipeline fails, runtime logs when a service misbehaves, quoted into the conversation.
What the review found. Severity, file locations and a fix prompt for the open branch.
How far the agent got. Ask about a session, read its handoff brief, collect the pull request.
The gate does not move#
Reaching your codebase from a phone does not widen what an agent may do with it. Research sessions are read-only. Coder sessions end where they always did, at a branch and a pull request with a preview environment and a review. Deploys and env var writes can each be set to ask, and the approval arrives in the chat.
Set it up once#
The connector is added and authorized in Claude Desktop or the web app, then it is on your account everywhere. It takes about a minute: connect Rock8Cloud to Claude.
Then open Claude on your phone and ask your codebase something.
Start free#
Put your codebase in your pocket. 60 days free, no credit card required.