MCP

kprompt mcp serve — read/plan-only Model Context Protocol tools for IDE assistants. Mutations return PlanResult JSON and never auto-apply.

kprompt mcp serve exposes kprompt to IDE assistants (Cursor, Claude Desktop, Windsurf, …) as a read/plan-only Model Context Protocol tool provider. kprompt here is an MCP tool provider, not an agent platform. Mutating prompts return a typed PlanResult and stop — the MCP surface never applies a change to the cluster. Approval stays a human action you run yourself (kprompt "…" --approve in your own terminal).

Transport

Newline-delimited JSON-RPC 2.0 over stdio. Your editor spawns kprompt mcp serve and talks to it on stdin/stdout; kprompt’s human-readable output goes to stderr, keeping stdout clean for the protocol.

kprompt mcp serve

Tools

Tool	What it does	Mutates?
kprompt.read	Natural-language read against the active kubeconfig. Returns PlanResult JSON.	No
kprompt.investigate	Multi-hop root-cause walk for a target.	No
kprompt.why	Causal analysis of why a target is failing/pending/crashing.	No
kprompt.timeline	Chronology of what happened to a target.	No
kprompt.impact	Reverse dependencies / blast radius for a target.	No
kprompt.plan	Compile a mutation prompt into a typed PlanResult. Never applies. Wipe-class intents hard-deny.	No
kprompt.tools	List detected integrations as JSON.	No
kprompt.doctor	Read-only environment health report. Never prints API keys.	No

Reasoning tools need a configured LLM provider because they compile natural language into intent. They honor your kubeconfig RBAC — no cluster credentials leave your machine. kprompt.plan is the explicit mutation surface: it returns the plan an assistant can show you; applying it stays out-of-band. The MCP server exposes no approval path.

Bridge story: IDE plan → CLI apply

One contract, two surfaces — same typed PlanResult. Use Cursor / Claude Desktop MCP for read and plan; apply stays a human action in a terminal you control. Never ask the assistant to “just apply it” over MCP — there is no apply tool and no --approve on the wire (ADR-0024).

Handoff

Cursor / Claude Desktop
  └─ MCP tool  kprompt.plan / kprompt.investigate / …
       └─ PlanResult JSON (risk · actions · blast radius)
            └─ you review in the IDE
                 └─ apply out-of-band:
                      kprompt "scale api to 3" -n payments --approve
                      # or: kprompt history rerun N --approve

Do in the IDE (MCP): kprompt.read / investigate / why / timeline / impact / plan / doctor / tools
Do in the CLI: the same prompts without MCP, plus kprompt "…" --approve after you trust the shape
VS Code PlanResult extension uses the same handoff — review in IDE, approve via local CLI

Editor configuration

Cursor

~/.cursor/mcp.json or project .cursor/mcp.json

{
  "mcpServers": {
    "kprompt": {
      "command": "kprompt",
      "args": ["mcp", "serve"]
    }
  }
}

Claude Desktop

claude_desktop_config.json

{
  "mcpServers": {
    "kprompt": {
      "command": "kprompt",
      "args": ["mcp", "serve"]
    }
  }
}

Use an absolute path to the kprompt binary if it is not on the editor’s PATH.

Safety

No remote auto-apply — the server never executes a mutation; --approve is not reachable over MCP
Hard-denies intact — wipe-class / namespace-delete intents are refused regardless of caller
Local trust — stdio transport is scoped to the operator who launched the editor; there is no network listener by default

Quick manual check

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | kprompt mcp serve