@agentutility/mcp-agentops

Trace, eval, and permission checks for autonomous work.

Frontier probe for agents that already run multi-step work and need trace briefs, flaky-step detection, tool-call diffs, MCP permission review, red-team prompts, and cost-per-success scoring.

Pricing: pay-per-call in USDC on Base. No subscriptions, no API keys. See per-tool prices below.

Install — Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "agentutility-agentops": {
      "command": "npx",
      "args": ["-y", "@agentutility/mcp-agentops"],
      "env": { "X402_PRIVATE_KEY": "0xYOUR_PRIVATE_KEY_HEX" }
    }
  }
}

Restart Claude Desktop. 4 tools appear in the tool palette.

Install — Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "agentutility-agentops": {
      "command": "npx",
      "args": ["-y", "@agentutility/mcp-agentops"],
      "env": { "X402_PRIVATE_KEY": "0x..." }
    }
  }
}

Funding

Send any amount of USDC on Base mainnet to the address derived from your X402_PRIVATE_KEY. The MCP server uses it to pay for tool calls automatically.

USDC on Base contract: 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913

Tools (4)

| Tool | Description | |---|---| | agent-trace-brief | (0.02 USDC/call) Turns a raw agent execution trace into a readable agent trace summary: an ordered step-by-step account of what the agent did, where it broke, and why. Built for tool call debugging when a run fails silently or an agent loops without explanation: send the trace array or log and get back failure points plus a root-cause hypothesis grounded only in what the trace actually shows, with anything it can't determine labeled unknown instead of guessed. Useful for agent run analysis after an incident, before a retry, or as part of a CI check on agent behavior. Use it as an agent debugging API, a trace root cause tool, or a post-mortem generator for agent pipelines. | | eval-case-generate | (0.02 USDC/call) Eval case generator for an agent or LLM task: describe the task and get back a draft llm eval dataset of {input, expected, rubric} cases spanning typical, edge, and adversarial scenarios. Pass example cases to steer style, and set n for how many to generate (default 8, max 20). Every case is labeled a candidate needing human review, not validated ground truth: this drafts a first pass at agent test cases, it doesn't certify them. Use it as an agent regression tests starter, an eval dataset generator for a new task, or a quick way to seed a rubric-graded benchmark before a human tightens it. | | mcp-tool-risk-score | (0.02 USDC/call) Heuristic MCP tool risk review for an MCP tool manifest before you wire it into an agent. Send the tool definitions (name, description, inputSchema) plus optional deployment context and get back a mcp permission review with per-tool scores for permission scope, prompt-injection surface, and destructive-action risk, each with reasons, plus an overall tool security score. This is a heuristic pass over manifest text, not a formal security guarantee or penetration test, so treat low scores as 'looks lower-risk on paper', not 'certified safe'. Use it as an agent tool audit step in CI, an MCP catalog screening pass, or a quick second opinion before granting a new tool broad access. | | tool-call-diff | (0.008 USDC/call) Deterministic tool call diff between two agent run sequences. Send before and after arrays of tool calls (name/args or tool/arguments shape) and get back which calls were added, removed, or reordered, plus per-call argument drift with the exact changed argument paths from a deep compare. No LLM involved, so results are exact and reproducible: the same replay comparison always returns the same diff. Built for agent regression diff work: confirming a prompt or code change didn't silently alter which tools an agent calls or what it passes them. Use it as a tool call diff API, an agent action drift checker, or a replay comparison step in CI before shipping an agent change. |

How it works

1. Agent calls a tool (e.g. agent-trace-brief).
2. MCP server POSTs to https://x402.agentutility.ai/agent-trace-brief.
3. The endpoint responds HTTP 402 with payment instructions.
4. The MCP server signs an EIP-3009 USDC transfer authorization with X402_PRIVATE_KEY and retries.
5. CDP facilitator settles on Base.
6. The endpoint returns the actual response.

The agent never sees the payment flow — it just gets the result.

Links

• Cluster overview: https://agentutility.ai/agentops/
• All MCP packages: https://mcp.agentutility.ai/
• Source: https://github.com/rooz21/x402/tree/main/packages/mcp-agentops

Version: 0.1.1 · License: MIT