Skip to content

MCP platform

Bounded, read-only fact-gathering tools that let local AI agents inspect the cluster without handing them an unrestricted kubectl or Grafana session.

AI & Agents In progress, first implementation running datahub-local-ai

A dark docking rack of six frosted apertures with light entering only

Problem

An agent is only as trustworthy as the facts it gathers. The easy path, giving a model API keys to Prometheus, Loki, and the Kubernetes API, fails in practice. A small local model with a large, similarly named tool surface makes unreliable choices, and a mutating tool one hallucination away from the cluster is a safety problem, not a convenience.

The platform already had the data: metrics in Prometheus, logs in Loki, workload state in the Kubernetes API, and tables in Trino. What it lacked was a narrow, explicit contract between that data and the agents.

Approach

A dedicated MCP (Model Context Protocol) server exposes a small set of purpose-named tools, one server per data source, and agents are restricted to explicit allowlists rather than a general-purpose administrator surface.

  • Each server answers one kind of question: query metrics, search logs, inspect workloads, or run SQL. Tool names describe intent rather than API shape.
  • The operations path is read-only. Any mutating capability is a separate policy decision, not an inferred action.
  • Tool names and arguments are verified against the live server, because a plausible-looking tool that returns nothing is worse than a missing one.

Technologies

  • Model Context Protocol server implementation in datahub-local-ai
  • Prometheus and Loki for metrics and logs
  • Kubernetes API with bounded, allowlisted operations
  • Trino, Polaris, and Garage for lakehouse inspection
  • Sympozium as the agent runtime that consumes the tools

Outcome

The MCP servers give agents a verifiable evidence path instead of open cluster access. They are the fact-gathering layer beneath the SRE agents: agents ask one bounded question, collect evidence from named tools, and report, with missing data surfaced as uncertainty rather than converted into a healthy result. The next step recorded on the landing is wider tool coverage with every mutation gated behind explicit policy and approval.

Further reading

Alvaro Santos Andres