org.nsgoods/nsgoods-workbench-mcp

nsgoods Workbench MCP

Read only index of the weekly x402 catalogue scan. Which endpoints are payable, what they cost, which hosts are gone.
No wallet. No sign in. No payments handled by this server.

What it answers

A hosted Model Context Protocol server over our weekly full scan of the x402 catalogue: is this endpoint payable, what does it cost, which host is gone, which endpoints match a keyword, and whether a signed nsgoods response verifies offline.

Hosted endpoint: https://mcp.nsgoods.org/mcp (Streamable HTTP, no sign in, 300 tool calls per IP per day). Landing and docs: https://mcp.nsgoods.org/ · Health: https://mcp.nsgoods.org/health

Data is point in time from the last full scan. Every answer carries as_of and the scan id. For a live check of one endpoint before paying, use the paid /payable oracle on the hub. This server is the index.

Connect

Claude (web or desktop)

1. Settings, then Connectors, then Add custom connector
2. Name: nsgoods. URL: https://mcp.nsgoods.org/mcp. No sign in.
3. New conversation, then ask for example: Using nsgoods, find the cheapest payable endpoint for sanctions screening on Base

Cursor: add to mcp.json

{"mcpServers":{"nsgoods":{"url":"https://mcp.nsgoods.org/mcp"}}}

Smithery: https://smithery.ai/servers/nikosble1926/nsgoods-workbench

Any MCP client that supports Streamable HTTP works the same way.

Tools

All tools are read only and idempotent (declared through MCP tool annotations).

Verdicts: PAYABLE, NOT_PAYABLE, MALFORMED_402, CANNOT_PAY_EITHER_VERB, UNREACHABLE, SKIPPED, UNKNOWN.

Example

A live find_endpoints("erc20-balance", 2) call, trimmed to two rows:

{
  "query": "erc20-balance",
  "total_matches": 6,
  "returned": 2,
  "endpoints": [
    {
      "resource": "https://api.onesource.io/api/chain/erc20-balance",
      "host": "api.onesource.io",
      "last_verdict": "PAYABLE",
      "host_gone_since": null,
      "price": "0.003 USDC on eip155:8453"
    },
    {
      "resource": "https://app.heinrichstech.com/v1/cdp/chain/erc20-balance",
      "host": "app.heinrichstech.com",
      "last_verdict": "PAYABLE",
      "host_gone_since": null,
      "price": "0.003 USDC on eip155:8453"
    }
  ],
  "as_of": "2026-09-13",
  "index_scan_id": "scan-20260913T050332Z"
}

Data and freshness

Index is rebuilt after each weekly full scan (14,652 resources, 1,971 hosts, 13,146 payable as of 13 September 2026). Prices come from a sweep of the 402 challenges (10,958 endpoints priced, median 0.01 USDC, 16 September 2026). A per endpoint sample with a GET only re-probe outcome per row is published under CC BY 4.0: https://x402.nsgoods.org/proof/payability-sample-2026-09-13.json

Resources that are no longer in the latest full CDP Bazaar scan are re-probed weekly in a separate reprobe scan: only non-PAYABLE rows, oldest first, up to 1,500 per week. Those rows keep in_latest_scan=false and carry their own last_checked_at and verdict_since; the latest full scan numbers (catalogue_stats, /health) do not include them.

Run your own

python -m venv venv && venv/bin/pip install -r requirements.txt
````
Put your own scans.jsonl (one JSON line per probe, see build_index.py for the fields) next to the scripts, then:
````
venv/bin/python build_index.py
venv/bin/python server.py   # listens on 127.0.0.1:4036, path /mcp
````
Docker:
````
docker build -t nsgoods-workbench-mcp . && docker run -p 4036:4036 nsgoods-workbench-mcp
````
Docker image is untested (built without a Docker host). Set WORKBENCH_HOST=0.0.0.0 inside containers. Feedback welcome.

Environment variables:

| var | default | meaning |
| --- | --- | --- |
| WORKBENCH_DIR | . | base directory for the index and data files |
| WORKBENCH_DATA_DIR | WORKBENCH_DIR | directory for scans.jsonl and watch state |
| WORKBENCH_HOST | 127.0.0.1 | bind address |

The server creates an empty index on first run if none is present, so tools/list works before you load any data.

## Privacy and logging

This server needs no wallet and no sign in. It keeps a minimal record of tool usage so we can operate and rate limit the service.

- Per tool call we store: timestamp, tool name, the names and lengths of the arguments, and ip_h, a daily keyed hash of the caller IP. We do not store argument values, and we do not store the raw IP or the MCP session id.
- When a client connects we store: timestamp, ip_h, user agent, the client name and version it reports, and the protocol version it requests. When a request to /mcp fails we store: timestamp, ip_h, user agent, HTTP method and status, the JSON-RPC method name, whether a session id was sent, the MCP-Protocol-Version, Accept and Content-Type headers, and the first 160 characters of the error. Never params or arguments.
- ip_h is HMAC-SHA256(daily_secret, ip) truncated to 16 hex characters. The secret is random, held only in memory, replaced at each UTC day change and never written down, so after the day ends the hash cannot be linked back to an IP.
- The same ip_h drives the free limit of 300 tool calls per caller per day. If the server restarts during a UTC day, a new secret is generated, so the counter restarts for everyone for the rest of that day.
- Records are deleted after 30 days. Records written before 26 September 2026 used a fuller format (caller IP, session id and the first 80 characters of each argument); they are deleted on the same schedule, the last of them by 26 October 2026.

The current field lists and retention are also reported live at GET /health under tool_log and client_log.

## Verification
Signed nsgoods responses use EIP-191 over canonical JSON (sorted keys, compact separators, ASCII escaped, signature and signed_by removed before hashing). The verify_signature tool checks the recovered address against the signers in https://x402.nsgoods.org/proof/index.json

Verifying what runs. `python3 scripts/pin_verify_signature.py server.py` prints the line range, byte range and sha256 of the verify_signature function body (def to end, decorators excluded, LF line endings, no trailing newline) and the sha256 of the whole file. A matching hash is repo evidence only: it shows that the published source is the one described. It is not deployment evidence. The stronger evidence for what the server runs is behavioural: calling the live tools and checking that the results match what this source produces.

## Links
Landing https://mcp.nsgoods.org/ · Hub https://x402.nsgoods.org · MCP section https://x402.nsgoods.org/#mcp
Weekly report https://x402.nsgoods.org/proof/payability-report-2026-09-13.html
Methodology https://x402.nsgoods.org/ata-audit/payability_index.json (method field)

MIT license.