Kubernetes creators open-source Mecatl, an agent harness built for the cloud

A coding agent today is usually a program that runs on a laptop: Claude Code began as a command-line tool, Codex and GitHub Copilot still rely on the local machine for most of their work. Latent Space reports that two of the three creators of Kubernetes, Craig McLuckie and Joe Beda, think that is the wrong shape for enterprise work, and have open-sourced Mecatl, an agent harness built from scratch to run inside a Kubernetes cluster.

Mecatl is the main product of Stacklok, which McLuckie and Beda founded. The repository, stacklok/mecatl, carries 226 stars at the time of writing and started on GitHub in late May. The project is open source under the same model as Stacklok's earlier MCP platform, ToolHive. Stacklok raised a 17.5 million dollar Series A in 2023 from Accel, Madrona and Bain Capital, Latent Space reports, and has recently pivoted from software supply-chain security into agent infrastructure.

What Mecatl changes about a harness

Beda told Latent Space that in a desktop-first harness, the LLM loop, the tool execution and the session state begin life as one process on one machine. Mecatl pulls those apart. The loop runs as its own cloud application. Tool calling and shell commands sit behind a separate boundary. Session memory moves into a store the team can back up and govern, rather than sitting as JSONL files on a developer's disk.

Beda said the "lift and shift" approach, dropping a desktop harness into a VM, breaks on the lifecycle of an agent and on being able to quiesce it: safely pause the agent while it waits for a human to answer. The Kubernetes analogy Beda and McLuckie reach for is the same control-loop idea that lets Kubernetes recover applications from failure, keep them running and scale them when needed.

Why this is being built now

McLuckie told Latent Space the pitch is "moving the locus of value from the desktop to the cloud." Code and context, in his words, "that's where the IP sits," and leaving it on a developer's machine makes it harder to manage. Beda said many agent tools were designed for small teams and do not fit a company trying to deploy an agent for "an engineering task force in the thousands or tens of thousands."

The company's business model is the "enterprise spine": consistent identity, authorization, policy and auditing across Mecatl and ToolHive. Stacklok's AI Gateway, the only product Beda said is not yet open source, handles access control, budgets, reporting and provider routing. It does not currently choose a model based on the task; McLuckie said that choice belongs in the harness, where more context is available.

The claim worth testing for your own team is Beda's about customization. He said Mecatl is designed to extend beyond a single-user, local harness, which also means the surface a platform team has to run goes up: a cluster to manage, a state store, a separate execution environment. The payoff is a harness an enterprise can govern the way it governs email, in Beda's own analogy. For a team already running coding agents at scale on laptops, Mecatl is the first open-source harness that treats the loop, the execution and the state as separable parts.

Source

Primary source: stacklok/mecatl on GitHub, Apache 2.0. Report: Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop? on Latent Space, by Richard MacManus.