GArness / Gelhaus Agent Harness
A control plane for AI coding agents that outlive their sessions.
A self-hosted web control plane for persistent and temporary AI coding agents across many repositories. A logical agent keeps its role, its memory and its task while the Claude Code session underneath it is handed off and replaced.
No companion
GArness is available only self-hosted: you run the control plane on your own server and the runner on your own machine. There is no hosted version.
An agent is not a session
GArness separates the agent from the process running it. A logical agent, such as the backend agent of a project, has a role, a memory, policies and a current task, and it survives indefinitely. The Claude Code session underneath it is replaced as often as it has to be, and the agent carries on as if nothing happened.
Handoff before the context runs out
GArness measures every session's context as it grows. At 60 percent the agent is told to finish its current task and start nothing new, at 70 it is reminded, and at 80 it writes a structured handoff. A fresh session picks up from that handoff in the same worktree. Compaction is treated as an alarm, never as the plan.
The operator stays in charge
Everything that needs a person arrives in one queue: decisions, approvals, halts and risks, split into what is blocking an agent and what is not. Tool calls pass through policies and, where they say so, wait for an approval. Agents reach tools through an MCP gateway and never hold a credential themselves.
Built on what exists
Plane is the record of the work, MCP carries the tools, A2A lets agents talk to each other, and every agent works in its own Git worktree, so the operator's own checkout is never touched. Durable work runs on Temporal, and events travel over NATS.
Self-hosted only
GArness runs on your own infrastructure: the control plane on a server you choose, and the runner on the machine where your agents work, dialling out to the control plane so nothing has to reach in.