bugtracker-mcp

Public, no-authentication MCP server over a bilingual (Hebrew / English) issue tracker. Built as a portfolio demo: anyone can add the URL to Claude or ChatGPT and use it immediately.

Live: https://mcp.raceus.co.il/mcp · landing page at https://raceus.co.il

Worker    bugtracker-mcp        (Cloudflare Workers, free plan)
Storage   D1 "bugtracker"       (SQLite, FTS5)
Cron      0 21,22 * * *         guarded to local midnight, Asia/Jerusalem

Running your own

cp wrangler.toml.example wrangler.toml
npx wrangler d1 create bugtracker        # put the returned id in wrangler.toml
npm install --legacy-peer-deps
npm run schema                            # applies src/db/schema.sql (destructive)
npx wrangler secret put RESEED_TOKEN
npm run deploy
curl -X POST https://<your-worker>/admin/reseed -H "x-reseed-token: $TOKEN"
./verify.sh https://<your-worker>

wrangler.toml is gitignored so deployments stay bound to whoever cloned it.

Surface

All three MCP primitives, where most public servers implement only tools.

Tools (11)

Resources — bug://{id} (templated, with listing) and bug://recent Prompts — triage_bug, weekly_summary

Related MCP server: FogBugz

Hebrew search

The interesting part. Hebrew attaches grammatical particles as prefixes (ה, ו, ב, ל, מ, כ, ש), so "export" appears in text as הייצוא and a user searching ייצוא gets nothing. FTS5 prefix matching (token*) only extends rightward and FTS5 ships no Hebrew stemmer, so neither helps.

Fix, in src/mcp/hebrew.js: normalise at index time by also indexing a particle-stripped variant of every Hebrew token, and OR-expand query tokens the same way. Matching becomes symmetric — ייצוא and הייצוא return identical results.

Deliberately shallow morphology: handles prefix particles, not roots, plurals or possessive suffixes. That covers the overwhelming majority of real misses.

Nightly reset

Cloudflare cron fires in UTC only. Israeli midnight is 21:00 or 22:00 UTC depending on DST, so both are registered and isLocalMidnight() guards the handler to fire on exactly one of them. Verified against summer and winter dates.

The reseed runs in scheduled(), never in a request path: the free plan allows ~10ms CPU per fetch() and rebuilding ~275 rows exceeds that comfortably.

Data

70 synthetic reports (38 Hebrew / 32 English) about a fictional generic SaaS app, plus 36 comments, 155 events and 14 photo attachments. Nothing real, no company named anywhere. Timestamps are stored as day offsets and materialised at reseed time, so the data always reads as recent.

Screenshots are drawn PNGs bundled with the Worker (src/data/photos.js) rather than object storage — R2 is not enabled on this account and the demo does not need it.

Commands

npm install --legacy-peer-deps   # agents@0.16.0 has a peer conflict; this is required
npm run schema                   # apply schema.sql to remote D1 (destructive)
npm run deploy
./verify.sh                      # 28 assertions against the live endpoint
python3 scripts/gen_seed.py      # regenerate seed data
python3 scripts/gen_photos.py    # regenerate bundled screenshots

Manual reseed (needs the RESEED_TOKEN secret):

curl -X POST https://mcp.raceus.co.il/admin/reseed -H "x-reseed-token: $TOKEN"

Gotchas

• npm install fails without --legacy-peer-deps. agents@0.16.0 declares a peer on ai@^6 while its optional @cloudflare/ai-chat peer wants ^6 || ^7.

• Reseed immediately after deploy can race propagation. A reseed run seconds after wrangler deploy executed the previous bundle and silently wrote unexpanded search rows. Wait, or verify SELECT MAX(length(explanation)) FROM bug_search is ~268 not ~198.

• raceus.co.il also hosts chat. (claude-chat-bridge) and tasks. (sukartask). The apex is a Pages project named raceus, not this Worker.