Everything tiny does, one static page per topic.
Install — Install the tiny CLI with Homebrew or a release binary, then one wizard sets up the cluster, the runtime and the agent token. No pods until your first session.
Try it in 5 minutes — A full run on a throwaway kind cluster: install tiny, start a Claude Code session, kill its pod and watch it resume. No real cluster needed.
Your first session — Start a Claude Code or Codex session on your cluster, close the laptop, and come back to finished work, a transcript and a branch.
Handoff — Move the local session you're in — files, dirty changes, transcript — to the cluster mid-conversation.
The fleet screen — One screen for the namespace, showing who runs and who is waiting on you.
Issues to PRs — The GitHub wiring, copy-paste ready: label an issue, get back a pull request.
The outbox — How finished work leaves the cluster when agents hold no credentials: a git bundle in the outbox, pushed by a separate short-lived job.
The gate — Risky actions park as Kubernetes objects until a human answers, and the answer runs with that human's credentials, not the agent's.
Claude & Codex — Claude Code and Codex, picked per session, signed in with the plan you already pay for.
Codex errors: active writer, code-mode-host — Why Codex says a thread already has an active writer (code -32600) after a crash or restart, how to clear it safely, and why tool calls fail without codex-code-mode-host.
Custom images — Run sessions in any glibc image you name, under a restricted container profile, or sandboxed with gVisor or Kata on GKE, AKS, k3s and more.
Spawning sessions — Sessions can start specialist sessions of their own. Each spawn parks as a question and only happens when a human approves it.
Messages, broadcast & uploads — Send messages into a session's durable inbox, broadcast to the whole fleet, and drag files into a workspace from your terminal.
Registry cache — A pull-through image cache in your namespace: one Docker Hub pull per image per namespace, 191 seconds cold and 9 seconds warm.
Artifact store — An S3-compatible MinIO store beside the sessions, one checkbox away, for builds, screenshots and datasets too big for git.
GitHub runner — An in-cluster Actions runner so the loop's jobs run next to the fleet.
Web page — A read-only fleet page with each session's blast radius, one checkbox away.
Egress policy — Default-deny networking for session pods, and an honest account of what it does not stop.
Threat model — What a prompt-injected coding agent can and cannot do inside a tiny session, the credentials it never holds, and the holes we haven't closed.
FAQ — How tiny compares with hosted sandboxes, Anthropic's own options and Google AX, what it costs, which plans work, and what it does not protect against.
All commands — Every tiny command and flag: starting, pausing and deleting sessions, attaching, the GitHub loop, egress, profiles and upgrades.