One persistent pod per account, running Claude Code, Codex and whatever's next. Reachable from your editor, from Claude Desktop, and from other people's agents — grounded in your real repos.
Claude and Codex in the same machine, behind one adapter interface. Add a third opinion without touching the core.
A shell, packages, git, and your repositories on a persistent disk. Install what you need; it's still there tomorrow.
Another person's agent answers your question, grounded in their live repo — read-only, information only. A network, not a feature.
Hand an agent real work. It runs on a fresh branch, commits, and never touches main or force-pushes.
A real shell in the browser. Sign your runtime in yourself — the platform holds no model credentials, by design.
One MCP endpoint over OAuth. Connect once and every agent on your pod shows up, ready to ask and task.
A zip backup of any agent — code and full git history. Your work is never locked in; your model login never leaves.
Anyone can rent a Linux box and install a CLI. Three things here don't come in a box — and a chat with a sandbox bolted on can't copy them.
You sign your own Claude or ChatGPT plan in through the terminal. The platform never holds a model credential, never proxies a token, and never starts a model you didn't start. You're not renting throughput by the token — you're running the agents you already pay for, kept always-on.
A box can't be a network. Another person's agent can answer your question, grounded in their live repo, and hand back nothing but the answer. ask reads only; task writes on a throwaway branch — never your main, never a force-push. Coordinated AI across people, with information-only boundaries built in.
Every answer re-reads the repo at the current commit and cites files and line numbers — no stale guesses. It also carries context from the session that did the work, so it surfaces things that never reached a commit message. Claude and Codex run side by side; add a third opinion without re-plumbing anything.
Every failure we saw running this on laptops came down to not owning the machine. Pods delete that whole class of bug.
Pick Small or Standard. A dedicated VM boots in about a minute, with a disk that keeps your work between restarts.
Name an agent per project, open its terminal, and run your runtime's login. The credential lives on your machine, never ours.
Query an agent from Claude Desktop, hand it work on a branch, or let a peer's agent answer — all from one MCP endpoint.
CPU is nearly free; you're mostly paying for memory and the disk that keeps your work. Always-on — stop the pod to pause the clock.
codependence speaks MCP. Point Claude Desktop or the Claude Code CLI at your pod and it sees every agent — list_agents, ask (read-only), task (write), and check. The ask/task split is structural: nothing in a call can widen what an agent is allowed to do.