worldbank-procurement-mcp

A remote MCP (Model Context Protocol) server exposing the World Bank's free Procurement Notices API (search.worldbank.org/api/v2/procnotices) as tools an AI agent can call. No API key needed on the World Bank side.

What was broken in the previous version, and the fix

The first deploy returned -32603 Internal server error on every request, failing in ~1ms of CPU time -- too fast to have even reached a fetch call. Root cause, confirmed against @modelcontextprotocol/server's actual changelog:

1. inputSchema was a bare object ({ name: z.string() }) instead of a full Standard Schema instance (z.object({ name: z.string() })). Cloudflare's own docs example uses the bare-object form, but the SDK's real behavior (per its changelog, PR #1689 "Support Standard Schema for tool and prompt schemas") requires a complete schema object. The bare form throws the moment the tool is registered -- which matches the instant, pre-fetch failure exactly. Fixed: every inputSchema in src/index.ts now uses z.object({...}).
2. Zod was pinned to v3 (^3.23.8); the SDK requires Zod v4 for Standard Schema support. Fixed: zod is now pinned to ^4.4.3.
3. Versions were floating or nonexistent. agents was "latest" (drift risk), and @modelcontextprotocol/server@2.0.0 doesn't actually exist as a stable release -- the real latest is the alpha line. Fixed: both are now pinned to exact, verified-published versions: agents@0.17.4, @modelcontextprotocol/server@2.0.0-alpha.2.

Also added: a try/catch around the handler that logs the real JS error to Observability/wrangler tail before responding, and a friendly plain-text response at / instead of a bare 404, so future debugging is faster.

Honesty check: @modelcontextprotocol/server is still an alpha package ("expect breaking changes until v2 stabilizes" per its own README) and Cloudflare's createMcpHandler pattern is only days old. I've verified every version and API detail above against the SDK's real changelog and Cloudflare's current docs, and this should deploy cleanly -- but I can't execute or test it myself (no network access in my environment), so "perfectly" is my best verified effort, not a guarantee against a dependency shipping another breaking change after this was written.

Tools exposed

• search_wb_notices -- keyword + filter search (country, sector, notice type, procurement method), paginated.
• get_wb_notice -- full detail for one notice by ID.

Deploying from a computer

unzip worldbank-procurement-mcp.zip
cd worldbank-procurement-mcp
npm install
npm start          # local dev server + MCP Inspector to test first
npx wrangler login
npx wrangler deploy

Test locally before deploying:

npx @modelcontextprotocol/inspector@latest

Point it at http://localhost:8788/mcp, Connect, List Tools.

Deploying from your phone only

If you don't have a computer handy, this replaces the existing world-bank-mcp Worker's code without touching a terminal:

1. On github.com (mobile browser), open the repo backing your Worker.
2. Tap into package.json, tap the pencil (edit) icon, delete the contents, paste in the new package.json from this zip, commit to main.
3. Repeat for src/index.ts (the big one -- this is the actual fix), wrangler.jsonc, tsconfig.json, and .gitignore.
4. If your Worker is connected to this repo for auto-deploy (check Workers & Pages → world-bank-mcp → Settings → Build), each commit to main redeploys automatically -- watch Deployments for the new one to go live.
5. If it's not connected to a repo yet: Workers & Pages → world-bank-mcp → Settings → Build → Connect to Git, pick this repo, and it'll deploy from there going forward.

Either way, after it redeploys, go back to the MCP playground and tap Add on https://world-bank-mcp.maksgstaad.workers.dev/mcp again.

Notes

• No secrets required -- the World Bank API is open.
• compatibility_date in wrangler.jsonc is pinned to 2026-07-30; bump it periodically per Cloudflare's Workers changelog.
• If this still fails after redeploying, check Observability again -- the new try/catch means you should now see the actual error message and stack trace in the log entry, not just a bare error code. Paste that back and I'll fix the real cause rather than guessing again.