HARP Stage H simulation

MCP is the safe agent interface for HARP.

MCP lets Codex, Cursor, Claude Desktop, Cline, or another MCP client ask HARP for cited reliability context without giving the agent shell access, Git write access, or Kubernetes mutation powers.

FastMCP read-only tools stdio or HTTP transport human approval for changes
Tech stack

The MCP stack in this project.

Protocol layer

MCP uses JSON-RPC style messages between a client and server. HARP implements it with fastmcp.

Packagefastmcp
Serverharp/mcp/server.py

Runtime layer

The MCP server shares HARP's index runtime. It can use BM25, optional OpenRAG hybrid search, evidence rendering, and Prometheus telemetry.

SearchBM25 plus OpenRAG RRF
Evidenceread-only renderer

Transport layer

Local tools spawn HARP over stdio. Remote agents can use streamable HTTP mounted into the FastAPI app at /mcp.

Localpython3 -m harp.mcp.server
Remote/api/harp/mcp/
Components

Tools, resources, and prompts are different things.

MCP clientCodex, Cursor, Claude Desktop, Cline
TriggerUser asks a question, alert triage starts, or prompt is selected
HARP MCPFastMCP server, read-only boundary
ToolsModel-controlled calls: search, hybrid search, triage, stats, health
ResourcesApp-controlled reads: docs index, document files, evals, profile
PromptsUser-selected templates: triage alert, investigate symptom, review RCA
Trigger flow

How MCP is triggered.

MCP is not a daemon that randomly starts doing work. A client starts or connects to the server, then a user or model action causes a tool/resource/prompt request.

01User asks

"Investigate this alert" or "search HARP for node disk pressure."

02Client chooses MCP

The app sees the HARP MCP server is configured and available.

03Transport opens

Local stdio spawns Python, or HTTP connects to HARP API's /mcp.

04Tool/resource call

The client sends a structured request like harp_search_hybrid.

05Cited answer

HARP returns JSON or Markdown with source paths and safety notes.

Interactive simulation

Step through an MCP call.

Click each step to see what is happening on the wire and inside HARP.


        

      

    

  


  

    

      
What this means

      
The agent gets context, not power.

      

        

          
Good use

          
Let agents search HARP, retrieve cited runbooks, summarize evidence, compare RCAs, and propose safe next diagnostics.

        

        

          
Bad use

          
Do not use MCP to expose shell commands, Kubernetes mutation, Git writes, reindexing, secret reads, or remediation execution.

        

        

          
Why it matters

          
MCP makes HARP usable by many AI clients without rewriting custom integrations for each tool.