mcp-server.md

MCP server (planned)

Status: planned / design note. A future surface, not yet released. It would expose the existing engine over the Model Context Protocol — no new analysis logic.

Goal

Let AI agents and assistants call the toolkit as tools. An agent investigating an incident could ask the toolkit to analyze a log, explain an error, or validate a manifest, and receive structured, deterministic, read-only results to reason over — grounding the agent in an auditable knowledge base rather than hallucinating.

Proposed tools

Tool	Input	Output
analyze_log	content, optional technology/source_kind	AnalysisResult (JSON)
analyze_yaml	content	AnalysisResult
analyze_terraform	content	AnalysisResult
explain_error	error	ExplainResult
validate_manifest	content, optional technology/filename	ValidationResult
list_signatures	optional technology	catalog entries

These map one-to-one onto existing engine methods. Each returns the same models documented in Output format.

Why it fits

The engine is already a clean façade with structured Pydantic outputs — ideal MCP tool results. And the read-only guarantee is exactly what you want when an autonomous agent is involved: the toolkit can only read text and return guidance; it cannot run commands or mutate infrastructure. See Security.

Proposed configuration

The MCP server would honour the same environment configuration — offline by default, with optional enrichment behind a provider key. A safe default is to disable enrichment for agent contexts so results stay deterministic and reproducible.

Sketch

// Example MCP tool registration (illustrative)
{
  "name": "analyze_log",
  "description": "Read-only DevOps log analysis: ranked root causes, diagnostic commands, fixes.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "content": { "type": "string" },
      "technology": { "type": "string" }
    },
    "required": ["content"]
  }
}

Internally the handler simply calls AnalysisEngine().analyze_text(content, technology=...) and returns result.model_dump().

Feedback

Building agentic workflows and want this? Open an issue (see Contributing) or try the hosted AI incident assistant in the meantime.