Stacklok wants to move AI coding agents from the desktop into the cloud

Stacklok is building Mecatl, an open-source harness designed to run AI coding agents as cloud-managed infrastructure rather than as tightly coupled desktop tools. The pitch is aimed at organisations that want to govern agent sessions, execution and access centrally, instead of leaving them scattered across developers’ machines.

Watch Desk analysis

What happened

In an interview reported by Latent Space, Stacklok co-founders Craig McLuckie and Joe Beda describe Mecatl as a Kubernetes-based harness whose agent loop is independent of the client, model provider, state store and execution environment. Stacklok began Mecatl as an open-source project in June, the report says.

The architecture separates the agent loop from more sensitive operations such as tool calls and shell execution, and from session management and memory. Stacklok’s other open-source project, ToolHive, is for running and governing MCP servers. The company also offers a commercial control plane intended to bring identity, authorisation, policy and auditing across its projects. Its AI Gateway handles access controls, budgets, reporting and provider routing; the founders say it is not yet open source.

Why it matters

Moving an agent into a cloud environment could make it easier for a company to manage sessions, permissions and computing centrally, particularly when a team grows beyond a handful of developers. It also changes where work and context live: the founders argue that code and other valuable company information are often rooted in local environments that are harder to manage consistently.

This is not simply a desktop tool wrapped in a container, according to Beda. The proposed separation of the agent loop, execution and state is meant to let organisations operate those pieces as managed infrastructure. That is a meaningful architectural bet, though the account describes Stacklok’s approach and aims, not independently demonstrated results at enterprise scale.

Our read

The useful idea here is not “agents, but in Kubernetes” as a phrase to decorate a slide. It is treating an agent’s identity, permissions, execution and memory as components that operators can govern separately. For organisations already running Kubernetes, that could be a more natural fit than asking every developer to maintain a personal agent setup.

The trade-off is control: central management can help with oversight, but it also puts more of an organisation’s agent workflow under the platform’s rules. Stacklok’s case is worth following as a design direction, not yet as proof that cloud-native harnesses have solved reliability or security.

What to watch

• Whether Mecatl documents how its separation of agent execution, state and tool access works in practice.
• Whether Stacklok’s AI Gateway becomes open source, as the founders say it is on the roadmap.
• Whether the commercial control plane gains features beyond the open-source projects and shows it can handle larger deployments.

Discussion spark: For coding agents used at work, should the default be a centrally managed cloud harness, or should developers keep more control on their own machines?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.