Agent Experience

Public MCP setup

Find what gets in an agent’s way.

Connect the Agent Experience MCP, then ask your assistant to check a sanitized tool catalog for concrete contract defects.

Connect once

Endpoint: https://agentexperience.tech/api/mcp

Copy setup text, or open Manual setup if copy is unavailable.

Manual setup
Agent Experience MCP
Server: https://agentexperience.tech/api/mcp
Transport: Streamable HTTP. No API key.
Use audit_agent_path with a task and sanitized tool catalog. For a repo, inspect it locally and submit only public-safe definitions. Never send secrets.
Public HTTPS URL audits are not enabled.
Setup: https://agentexperience.tech/connect/

Run one audit

Available now: sanitized catalog audits. Public URL fetching is not enabled yet.

The tool supports a task and public HTTPS URL, a sanitized catalog, or both. URL checks require the server’s public-network safeguards to be enabled.

Example: “Use audit_agent_path to review this repository’s tool contracts. Read the catalog locally, remove secrets and private examples, then submit only the sanitized definitions.”

Illustrative report shape

This example shows the format, not a finding about any real site.

llms.txt serves an HTML fallback

  • Observed fact: /llms.txt returned HTTP 200 with text/html, and the body began with an HTML document instead of plain Markdown.
  • Fix: Serve the intended plain-text llms.txt content, or return a real 404 if this optional file is not published.
  • Retest: Run audit_agent_path again and confirm the content type and opening body match the intended plain-text resource or a clear missing-file response.
  • Human follow-up: If the result needs interpretation, contact Himadri Mishra.

What is checked

  • For a public HTTPS URL: the entry page, robots.txt, and llms.txt as up to three resources, with bounded redirects.
  • For a supplied sanitized catalog: deterministic contradictions in names, descriptions, and schemas.
  • The report separates observed facts, suggested fixes, and retest steps.

Limits and privacy

The audit does not run a browser, operate an agent, access private repositories, log in, invoke your tools, or predict success. Failed public access is reported as inconclusive, not as a product defect.

Submit only public URLs and sanitized catalogs. Do not include private code, customer data, tokens, or credentials. The endpoint may fetch the public URL you provide; it does not need credentials and does not change the target site. Application usage events record tool name, outcome, and duration, not your task, URL, or catalog.