A container that already knows your code.
Every workspace gets a persistent, self-hosted pod — bound to your GitHub provider, every repo cloned, env injected, toolchains ready. Open a terminal in the browser and run Claude Code, OpenCode, or hand it to Hermes. Every step it takes lands back on your timeline.
It boots into your whole world.
A pod registers with the platform, pulls a manifest, clones every repo onto a persistent volume, injects your env, and keeps a tmux session alive. No re-discovery, no "which directory am I in" — the agent gets its entire working world on the first call.
[pod-runner] registered · working_root=/workspace · desired=on [pod-runner] clone forked.online ✓ [pod-runner] clone lead-caller ✓ [pod-runner] clone drive-x · restaurant ✓ [pod-runner] reconciled: 6 cloned · env → /workspace/.pod.env [pod-runner] MCP configs synced · claude=1 opencode=1 [pod-runner] terminal server on 127.0.0.1:43187 [pod-runner] heartbeat · online
miseon-demand toolchains — mise install reads .tool-versions and pins the exact Ruby/Node/Rust per repo.agent CLIsClaude Code, OpenCode, and @forked-online/mcp pre-installed and auto-wired.native libsgit · tmux · ripgrep · libpq · libvips · openssl — the things real builds need.terminala truecolor xterm-256color tmux session, streamed to your browser.{
"pinned_memories": [ // ride along on every get_work_item
{ "kind": "guardrail", "title": "One commit per completed todo" },
{ "kind": "convention", "title": "Branch first — feat/wi-XXXX-slug" },
{ "kind": "convention", "title": "WI-XXXX in every commit + PR title" }
],
"recall": [ // the long tail, pulled on demand
{ "kind": "gotcha", "title": "No fixtures — remote PG lacks SUPERUSER",
"author": "deepseek", "recall_count": 7 }
]
}
Run the agent you love. In the browser.
The pod ships the real CLIs. Open a tab and you're in the actual TUI — truecolor themes and all — talking to your repos with the forked-online MCP already connected.
▄▖▄▖▄▖▖ ▖▄▖▄▖▄▖▄▖ ▌▌▙▌▙▖▛▌▌▌▌ ▌▌▌▌▙▖ ▙▌▌ ▙▖▌▌▙▌▙▖▙▌▙▌▙▖
Not at your desk? Hand off or Implement a work item and a managed agent runs in the pod — plans, builds, and won't ship until its gate passes: coverage · tests · self-review.
One canonical surface. Everything logged.
21 MCP tools — the same surface whether you drive Claude Code from a pod tab or dispatch Hermes from a work item. Read tools ground the agent; write tools land on your timeline.
get_work_itemverify_repolist_work_itemsget_projectlist_projectslist_todoslist_activitiesrecallcreate_work_itemupdate_work_itemset_work_item_statuscreate_todocomplete_todolog_activitypost_thread_messageassign_workitem_to_branchrequest_repo_accessrequest_clarificationrecord_discussionrememberforget- 01 checkpoint — read the timeline — never restart from zero
- 02 plan — propose, then wait for your go-ahead
- 03 break down — fire create_todo in parallel
- 04 implement — verify_repo, then commit per todo
- 05 gate — coverage · tests · self-review
- 06 ship — assign_workitem_to_branch → in_review
Stop renting
your workflow.
A command center for one operator — with a container the agents work inside, and a timeline that proves you built it.