Public MCP server

Let your assistant use these guides

MCP (Model Context Protocol) is the open standard that lets an assistant call tools. Connect this read-only server and your assistant can search and read the guides, look up terms, fetch the rubric, and check a tool list you give it for problems.

Endpoint

https://agentexperience.tech/api/mcp

Sign-in: NoneNo API key
Changes anything?: NoEvery tool is read-only
Tools: 51 audit, 4 reference
Status: Tool-list audits workChecking a public URL: not yet

Connect once

Paste this into your assistant’s MCP settings.

Agent Experience MCP
Server: https://agentexperience.tech/api/mcp
Transport: Streamable HTTP. No API key.
Use audit_agent_path with a task and a sanitised tool catalog. For a repo, read it locally and send only the tool definitions.
Public HTTPS URL audits are not enabled yet.
Setup: https://agentexperience.tech/connect/

Ask for an audit

audit_agent_path is the one executable check on this site today. Give it a sanitized list of tools and it checks duplicate names, missing descriptions, undefined required properties and example arguments against a supported schema subset.

Try this prompt

Use audit_agent_path to review this repository’s tool contracts. Read the catalogue locally, remove secrets and private examples, then submit only the sanitised definitions.

Checking a public web address (its entry page, robots.txt and llms.txt) is built, and switched off until the server’s public-network safeguards are on.

Read the report

Try this public-safe catalog input. The description uses order_id, but the closed schema defines orderId and requires order_id. The audit can detect that specific schema defect without fetching a URL or calling the tool.

{
  "task": "Check this fictional order-status tool's input contract.",
  "catalog": [
    {
      "name": "get_order_status",
      "description": "Read the status of one order using its order_id. Does not change the order.",
      "inputSchema": {
        "type": "object",
        "properties": {
          "orderId": {
            "type": "string"
          }
        },
        "required": [
          "order_id"
        ],
        "additionalProperties": false
      }
    }
  ]
}

Synthetic example. No product action is executed.

`order_id` is required but not defined

Evidence: /catalog/0/inputSchema/required/0 names order_id; the closed schema defines orderId.
Fix: If order_id is the intended public input, rename the property to order_id. Confirm the implementation uses the same name.
Check the fix: Submit the corrected catalog to audit_agent_path. The required-property finding should disappear. Test the actual tool separately; a clean static report does not prove it runs correctly.

Continue with the tool-description test cases to review nearest-tool selection, inputs and recovery. An optional workflow review is available if you want help applying the checks.

What the five tools do

Tool	Use it to	Not for
`audit_agent_path` The one executable check	Check a sanitized tool list for duplicate names, missing descriptions, undefined required fields and invalid examples in the supported schema subset.	Running a browser, logging in, calling your tools, or predicting success.
`search_guides` Reference	Find which of the 29 guides covers a topic.	Reading a guide (use `get_guide`) or defining one term (use `get_glossary_term`).
`get_guide` Reference	Read one guide in full, as Markdown, to quote or cite it.	Collecting the whole library; fetch llms-full.txt once instead.
`get_readiness_rubric` Reference	Get the Open Agent-Readiness Rubric as JSON to assess one job: a profile across six tracks, each rated at one of four levels.	Scanning a site for you, or explaining why a check matters (read the rubric guide).
`get_glossary_term` Reference	Get a short definition of a term such as AX, MCP or llms.txt.	“How do I…” questions; use `search_guides`.

Limits and privacy

For each call the server records the tool name, the outcome and how long it took, not your task, address or tool list. If it cannot reach a public page, the result says “inconclusive”, not “defect”.

Privacy · Terms of use · Support and limits