Agents
Queryable, not just executable
The engine's descriptions are data. That makes it something an agent can read, reason about and propose changes to — through the same door a human uses.
The Microcelium Substrate control plane exposes an MCP server: 108 tools — 98 read-only views over fleet, projects, machines, jobs and config; 10 that enqueue a job. Every one of the writers enqueues a typed, logged, replayable job for the runner to execute — none acts directly — and each tool gates itself on the caller's role, so the surface area is not the permission grant.
Ten of those tools sit on the render path this site is about: re-render a project, diff two contexts key by key, list the compiled bundles, map a rendered stack, pin a project to an engine revision. Each only means anything because the templates are the same everywhere and the render is the same bytes every time.
AI sessions propose changes to the engine itself; every one lands through a merge request that a named engineer reviews, rendered and diffed the same way as any other change. Every commit carries a session link.
Operator- or schedule-triggered. Nothing here acts on its own. Because every service describes itself, an assistant can answer questions about an estate and propose changes through the same reviewed route an engineer uses.
what an agent can ask
- How staging differs from production — aligned, key by key. Only possible because both are one set of templates with a different values file.
- What a rendered stack looks like — the compose graph of one project.
- Which engine revision a project is pinned to — the drift problem, made readable.
- Re-render it — as a job the caller's role permits, not a decision the agent takes.
micro‑glue is part of the Microcelium tooling that builds your application, and what it builds is yours to keep. If you run several products on machines you own and want them built the same way every time, that is the conversation.