waom-mcp

Cross-platform Ironbound WAOM. A Model Context Protocol server that exposes the autonomous-ops loop — monitor → decide → fix → verify — as four tools, for Claude Code, Codex, and Cursor.

🚀 Run it: npx @shizhanyu13/waom-mcp

Why this exists

The classic WAOM loop — Watch, Assess, Operate, Maintain — says: detect it, fix it, verify it. @shizhanyu13/dsh-waom does that loop inside the DeepSeek Harness, driving a DSH subagent to execute the fix. But that ties the loop to DSH.

@shizhanyu13/waom-mcp unlocks the same reasoning for any MCP-capable agent. It ports the pure decision/evaluation logic from dsh-waom and exposes it as four tools. You don't need DSH, and you don't need a subagent — the host agent (Claude Code, Codex, Cursor) is the executor. The MCP server is the brain; the host is the hands.

Division of labor

• waom-mcp (this package) — the senses + planner + verifier: it probes a target, decides whether a fix is needed, hands the host a fix instruction, and independently re-verifies.
• Host agent — the executor: it calls monitor → decide → fix → verify, and when fix returns an instruction, it applies the fix with its own tools, then confirms with verify.

Quickstart

Add to your MCP client config.

Claude Code (.mcp.json)

{
  "mcpServers": {
    "waom": {
      "command": "npx",
      "args": ["-y", "@shizhanyu13/waom-mcp"]
    }
  }
}

Codex (config.toml)

[mcp_servers.waom]
command = "npx"
args = ["-y", "@shizhanyu13/waom-mcp"]

Cursor

Settings → MCP → + Add new MCP server → name waom, command npx, args ["-y", "@shizhanyu13/waom-mcp"].

Tools

| tool | purpose | inputs | |---|---|---| | monitor | Probe an HTTP health endpoint (HEAD). Returns reachable + healthy. | id, url, healthyCode? | | decide | Turn a health snapshot into a fix decision (heuristic fallback). | healthy? or id+url | | fix | Build the fix instruction for the host agent to act on. | action, rootCause?, constraints? | | verify | Re-probe a target and independently evaluate it (GAN-style). | id, url, healthyCode? |

The happy-path gate is HTTP reachable + healthy code (default 200).

The loop (let the host drive it)

monitor(id, url)        → { reachable, healthy }
decide(healthy: false)  → { needs_fix: true, action: "fix", root_cause: "target unhealthy" }
fix(action, rootCause)  → "[WAOM fix] fix: target unhealthy\n..."
  … host applies the fix …
verify(id, url)         → { passed: true, reason: "target healthy after fix" }

This is deliberately four small tools, not one opaque run: the host keeps control of what to fix and how, while waom supplies the detection, decision, and independent confirmation.

Default behavior

• Inert until called — the server only responds to MCP tool calls. No background ticker, no surprises.
• Pure, DSH-free — 0 DeepSeek Harness deps. Works anywhere Node 20+ runs.

Relationship to @shizhanyu13/dsh-waom

dsh-waom is the same loop inside DSH (its fix executor is a DSH subagent). waom-mcp reuses the pure probe / heuristicDecide / evaluate / buildFixPrompt functions and swaps the execution layer for "the host agent". Upgrade dsh-waom's logic → the same decisions land in waom-mcp's tools.

License

MIT