AgentPod is a console for the machines my agents live on. A small Go node agent runs on each host, dials out to a hub over WSS, and detects the agent harnesses running there: Hermes, OpenClaw, Claude Code, Codex, OpenCode and Pi. From one console I can browse a station’s files, tail its logs, open a terminal, check its health and clean up after it.
The first version, in December 2025, was an app for running one coding agent in a Docker sandbox, with OpenCode wired in as the agent. By mid-2026 my own agents ran on laptops, a VPS and cloud sandboxes under several harnesses, and the problems I kept hitting were operational ones: a full disk, a config file that kept getting overwritten. So I rebuilt AgentPod as a fleet console and froze the old product as a git tag. I wrote about that in The agent platform I kept rebuilding.
A few decisions shape it:
- Attach first. It adopts runtimes you already run, and can also provision new ones.
- Nodes dial out. No inbound ports, so a laptop behind NAT joins on the same terms as a VPS.
- One job. AgentPod does not orchestrate work or chat. SuperPipeline owns the work and SuperMessage owns the conversation. Principals and dispatch grants tie runtime operations to an agent’s identity.
- Deliberate updates. Node agents never upgrade on a timer; I update one node or the fleet on purpose.
It is in active development. Node-agent binaries are published for Linux and macOS, while the hub and console run from source. It currently serves a single bootstrap tenant rather than independently managed organizations.