agentleFS
DaveGoldDaveGold/mcp-metadata-demo/.claude/skills/mcp-compat-check/SKILL.md

mcp-compat-check

Skill7 starsChanged 21 days agoMITView on GitHub

Checks in about a minute whether an MCP server's metadata survives the clients it will run in (Claude Code, claude.ai, Cowork, ChatGPT Chat and Work, Codex): reads tools/list and the server instructions over HTTP or stdio, and scores each tool against each client's measured limits on descriptions, server instructions, input schemas and output schemas. Invoke when asked to check, health-check, lint or score an MCP server, to ask "will this server work in ChatGPT / Codex / claude.ai", "why does the agent ignore my description or parameter docs", or to compare public servers. Hands fixes to the rich-domain-mcp-server skill.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never put a token in the command yourself: ask the user to export it as an environment variable, and pass the variable.

    line 33
  2. Very specific:

    Start with the "Start here" list.

    line 74

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

mcp-compat-check — does the metadata reach the model on every client?

A server can ship perfect metadata and still not deliver it, because each client cuts or drops different surfaces. This check compares what a server ships with what each client delivered when it was measured (2026-09-27; evidence: rich-domain-mcp-server → references/evidence.md, tag HD). It is static: no model and no tool calls, so it runs in seconds and gives the same answer every time.

Run it

node .claude/skills/mcp-compat-check/check.mjs https://example.com/mcp
node .claude/skills/mcp-compat-check/check.mjs --header "Authorization: Bearer $TOKEN" https://example.com/mcp
node .claude/skills/mcp-compat-check/check.mjs --stdio -- node dist/stdio.js
node .claude/skills/mcp-compat-check/check.mjs --json https://example.com/mcp   # for CI or diffing
node .claude/skills/mcp-compat-check/check.mjs --oauth https://mcp.example.com/mcp  # OAuth servers (needs @modelcontextprotocol/sdk)

Never put a token in the command yourself: ask the user to export it as an environment variable, and pass the variable. For OAuth servers use --oauth: it opens the user's browser, the user signs in and approves, and the token lives only in the checking process. Run one OAuth server at a time, and tell the user which service the next tab is for.

What it checks, per client column

surfaceClaude only (Code, claude.ai, Cowork)ChatGPT / Codex onlyall clients
tool description≤ 4,096 (Claude Code 2.1.296+ for a tool sent up front, and Cowork; 2,048 before 2.1.296)no cut measured≤ 4,096
server instructions≤ 4,096, and nothing required (claude.ai chat gets none)≤ 512, and nothing required (Work gets none)≤ 512, nothing required
input schemawhole≤ 5,000 serialized, or every describe is dropped (bisected on Codex; ChatGPT Work consistent, not bisected)≤ 5,000
output schemanever delivered: meaning there is review-onlydelivered on Codex and Workreview-only

It also flags: guidance that lives mostly in the server instructions, tools with a first sentence too short to be found by clients that load tools by search, missing annotations, and the size of tools/list, which is re-sent every turn.

Reading the report

The Verdict table has one column per client group and one row per surface:

rowwhat it counts
overallthe worst surface in that column, which surfaces fail, and how many tools are affected
tool descriptiontools whose description is longer than the column's cut
server instructionstheir length against the column's budget, and whether the guidance lives there; within budget is green, with a note that some clients deliver none
input schematools whose serialized schema is over 5,000, so Codex and ChatGPT Work drop every describe
output schematools that ship one; on the Claude clients it never arrives, so meaning found only there is flagged
  • 🟢 portable: within every measured limit for that column.
  • 🟡 host-dependent: it arrives on some clients, or is never needed; review it. Output-schema meaning is fine if the same meaning travels in the field names or the response, which a static check cannot see.
  • 🔴 not portable: the server relies on a surface or size that some measured client in that column does not deliver.

A portability finding, not a quality rating. A protocol-correct, well-built server can be red: the protocol lets it ship metadata that some clients never pass on. The report's last section lists the rule behind every cell, so anyone can recompute the matrix.

Start with the "Start here" list. For each fix, use the rich-domain-mcp-server skill: its Budgets by target client table has the rule and its evidence, and its REQUIRED guidance-tool pattern holds detail that does not fit a budget.

Limits

  • One day's client versions. Re-measure before relying on a number: the quote probe and the Codex request trace are in rich-domain-mcp-server → references/delivery.md.
  • It does not call tools, so it cannot see the response, where most meaning should live.
  • A green report means the metadata arrives. It says nothing about whether the metadata is right.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.