gold402-mcp

An MCP server over gold-402 — the curated x402 directory maintained by 24K Labs. Gives any MCP-capable agent (Claude, Cursor, or anything speaking MCP) four tools for finding x402 services and for checking an endpoint before it spends money against it.

Status: v0.1 — tested, packaged, not yet published. Runs over stdio.

Tools

Free vs paid

Four of the five tools are free and stay free — search, shelves, entry lookup, and the live probe. No key, no wallet, no account.

gold402_verdict is paid: $2.00 USDC on Base (the live endpoint states the exact price in its 402; the number here is as of 2026-09-02). It is the only one, and it buys the thing the free probe explicitly does not give you — a human judgement. A valid 402 proves the door works; the verdict is a maintainer saying it was worth a stranger's money.

This server never holds a wallet and never pays on your behalf. Call gold402_verdict once with no payment_signature and it hands back the decoded x402 terms — price, network, asset, payTo, timeout. Sign them with your own wallet and call again with the signature. The money goes from you to 24K Labs; this process is not in the path.

gold402_check_endpoint is the door check the maintainers run, made callable. It works on endpoints that are not in the directory, which is the point: an agent about to pay a stranger can ask first.

What the answers claim, and what they don't

The shelf is admitted under two rules and the difference is load-bearing. A service is listed only if a maintainer probed its endpoint and it answered an x402 request correctly at the time of review. A library, framework, guide or community resource has no payable endpoint, is never asked for a 402, and is listed on public reachability alone — do not read one of those as evidence that it speaks the protocol. Either way we re-probe periodically. It is not an audit of the provider, a guarantee of uptime, or a promise that any call will succeed. Which entries carry a dated receipt — and which are unproven — is published at https://24klabs.ai/.well-known/touchstone. A valid 402 proves the door works — not that the service delivers. Every response carries an as_of stamp and the upstream it came from.

Absence from the directory is not a verdict against a service. It's a curated list, not a census.

Install

Works today, nothing to sign up for:

pip install git+https://github.com/Haustorium12/gold402-mcp.git
gold402-mcp

Or with no install step at all:

uvx --from git+https://github.com/Haustorium12/gold402-mcp.git gold402-mcp

From a clone:

python3 -m venv .venv && ./.venv/bin/pip install -e .
./.venv/bin/gold402-mcp

Not on PyPI yet, so plain pip install gold402-mcp will not resolve. The git form above is the real install and is verified working.

Client config

Claude Desktop / Claude Code (claude_desktop_config.json or .mcp.json):

{
  "mcpServers": {
    "gold402": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Haustorium12/gold402-mcp.git", "gold402-mcp"]
    }
  }
}

Installed with pip? Use "command": "gold402-mcp" with no args.

Data source

Two upstreams, both public, no key, no auth, no scraping:

• https://24klabs.ai/api/directory/search?q=<text>&section=<id> — full-shelf search, free and ungated. gold402_search and gold402_get_entry query this directly, so a search runs over every curated entry, and each result carries the date it was last knocked.

• https://24klabs.ai/directory.json — a summary view: shelf structure, true per-shelf counts, and the 25 most recently listed entries. Used for gold402_list_sections and as the fallback if search is unreachable. Cached 15 minutes; a stale cached copy is served with its age disclosed rather than hidden.

Coverage, stated plainly. Every response reports the curated total and how many entries the query actually ran over. If the search endpoint is down and this server falls back to the summary, the answer says so in its coverage line and a miss is reported as unknown, never as absent. gold402_check_endpoint works on any endpoint whether or not it is listed.

Until 2026-09-02, search ran over the 25-entry summary only and reported that as the whole shelf — 4.9% coverage presented as 100%. Fixed by querying the search endpoint; the coverage line is now the shelf's own number.

Tests

./.venv/bin/python test_tools.py    # tool-level, incl. error paths
./.venv/bin/python test_stdio.py    # end-to-end over the real MCP protocol

Both pass as of 2026-08-20, including against the installed console script (clean venv, pip install from the built wheel) — not just the source file.

Curated by 24K Labs · directory: https://github.com/Haustorium12/gold-402