Three ways AI shows up. One platform underneath.
An agent that investigates, a governed door for the agents you already use, and observability for the AI you ship. All on the same lake, under the same RBAC, and in all three the engineer stays the one who decides. A context-aware assistant inside the product follows.
You set the guardrails. It stays inside them.
Dexter, the Kloudfuse AI SRE, onboards the way a new engineer does, learning your services, dependencies and incident history. It investigates across everything inside your boundary and hands back a drafted RCA with the evidence attached. A proposal, not a change.
- It reads more than we collect. The lake is primary, but it pulls from Datadog, Grafana and New Relic too, so an investigation is never limited to what you have already migrated.
- The reasoning stays inside the perimeter, and so does the bill. Dexter runs in your VPC on your own model, with an outside provider only if you want one. No AI surcharge, and no token cost that scales with how bad your quarter was.
- It drafts, it does not fire. Every report covers what happened, blast radius, timeline and a recommended fix it will not execute without your say-so. GitHub, Jira and Slack connect only behind a human approval.
- The engineer's call wins. Fork your own line of investigation and promote what holds up. Nothing Dexter concludes overrides the person working the incident.
Most MCP servers wrap an API. This one is built for how agents work.
Your engineers already have an AI assistant open. The Kloudfuse MCP server is the governed door between it and your production telemetry, whether that is Claude, Cursor, Codex or something your team wrote. One deployment serves the organization, and your data never moves to reach it.
- An agent does not need your whole platform to answer one question. Nine toolsets, chosen per connection: logs and alerts for on-call, traces and profiles for latency. A narrower surface means fewer wrong turns and a smaller token bill.
- Correlation happens in the server, not the context window. Ask about a service and get health, traces, logs and dependencies already correlated, rather than four raw result sets the model reassembles and charges you for twice.
- Query Safety Mode stops a runaway query before it reaches storage, and names the limit it hit so the agent rewrites rather than retries.
- The agent signs in as the engineer, through your own SSO, and sees exactly what that person sees. Every question and query lands in your lake, so the security review is one you can run yourself.
A model call is a span. An agent run is not.
The model call lands as a span in the same trace and the same lake as the API around it. But the call was never the hard part. Tracing was built for services calling services, and now agents call agents, where every step depends on what the last tool returned and no single step looks wrong alone. Agentic observability is tracing catching up with what it is now tracing.
- Today, the call is open to you. Prompt and completion text, tokens, latency and errors ride on the span, so a rate-limited provider and a slow model stop looking alike. LangChain and LlamaIndex instrumented out of the box.
- Token spend traces back to a cause. Usage arrives as Prometheus metrics split by service, model and provider, so a spike lines up against a deploy or a customer.
- Three people need the same record. Engineering triages, support explains, compliance audits. Build it for one and the other two get nothing, which makes this a product decision rather than a debugging feature.
- It belongs in the lake you already own. The alternative is every team designing its own capture format from scratch. Run history is trace data with a harder shape, and it should live under the retention you already set for audit.
