Agent Daily MCP Server

Agent Daily: is this tool up right now?

A free, hourly reliability index for the APIs and MCP servers AI agents call: 65 providers across LLMs, search, scraping, email, vector DBs, voice, payments, messaging, maps, CRM, storage and public remote MCP servers, with 7-day uptime, p50/p95 latency and incident history.

Built for agent builders. Check before your agent calls a tool, fall back when a tool is down, and stop burning retries on outages.

• Status page: https://gregpunch.github.io/agent-daily/
• Machine-readable: https://gregpunch.github.io/agent-daily/status.json
• Agent guide: https://gregpunch.github.io/agent-daily/llms.txt
• MCP server (no key): https://agent-daily.agentdaily.workers.dev/mcp, with tools list_tools, tool_status

claude mcp add --transport http agent-daily https://agent-daily.agentdaily.workers.dev/mcp

Consume status.json

import requests
s = requests.get("https://gregpunch.github.io/agent-daily/status.json", timeout=5).json()
tool = next(t for t in s["tools"] if t["name"] == "OpenAI API")
if not tool["up"] or tool["vendor_status"] in ("major", "critical"):
    ...  # route to a fallback provider

Fields: schema_version, generated_at (UTC), summary.{up,total}, and tools[] with name, category, up, http_status, latency_ms, vendor_status, uptime_24h_pct, checks_24h, error. Data is refreshed hourly; treat it as stale if older than about 2h. Please cache.

Limits: these are reachability checks only (an expected HTTP code), not correctness or quality tests. There is a single probe location.

A free public status page for 65 tools AI agents depend on. This is Phase 1, a two-week validation.

Run locally

python3 check.py   # probe all tools -> results.db (SQLite)
python3 build.py   # render site/index.html

Files: tools.json (config), tools.md (list + method), check.py, build.py, site/ (static output), preview.png.

Deploy plan (free tiers)

Recommended: GitHub Actions + GitHub Pages ($0).

1. Push this folder to a public GitHub repo.
2. Add .github/workflows/check.yml with schedule: cron: "0 * * * *" (hourly). Steps: checkout, pip install requests, python check.py && python build.py, commit results.db back (or store it as an artifact), then deploy site/ with actions/deploy-pages.
3. Turn on Pages (source: GitHub Actions). Public repos get unlimited Actions minutes.

Alternative: Cloudflare Worker cron + D1 + Pages ($0). Port check.py to a Worker scheduled() handler (free tier: 100k requests/day, cron triggers included). Store rows in D1 (free 5 GB), and render the page or serve JSON from the Worker or Pages. This option is the better base for the paid MCP/API later.

Estimated monthly cost

MCP server (mcp-worker/)

This is a stateless, read-only MCP server on a Cloudflare Worker. It uses Streamable HTTP at /mcp and offers two tools: list_tools(category?) and tool_status(name). Both read status.json with a 5-minute cache. It also serves /.well-known/mcp/server-card.json.

cd mcp-worker && npm install && npx wrangler dev   # local
npm test                                           # MCP SDK client check against localhost:8787
npx wrangler deploy                                # needs CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID

MCP tools (v0.2.0)

Remote endpoint: https://agent-daily.agentdaily.workers.dev/mcp (Streamable HTTP, no auth, read-only)

• recommend(task): maps a phrase or category ("send email", "web search", "llm", "scrape a page") to a category and returns the healthiest provider plus 1-2 fallbacks, with reasons and freshness.
• incidents(name?, days=7): outages and vendor incidents with UTC start/end.
• list_tools(category?): matches either category or routing task.
• tool_status(name): full record, including 7-day uptime and p50/p95.

Data: status.json (current + 7d stats) and history.json (incidents), both regenerated hourly.

LLM Live Probe (v0.3.0)

• Leaderboard: https://gregpunch.github.io/agent-daily/llm/
• Methods: https://gregpunch.github.io/agent-daily/llm/methods.html
• JSON: https://gregpunch.github.io/agent-daily/llm/leaderboard.json
• MCP: llm_leaderboard(sort?) at https://agent-daily.agentdaily.workers.dev/mcp

https://gregpunch.github.io/agent-daily/llm/: every 15 minutes ~20 popular models (via OpenRouter, plus optional direct OpenAI/Anthropic/Gemini control checks) get "Reply with exactly the word pong". Leaderboard (score = 60% 24h correctness + 40% speed), per-model detail pages and a methods page. Data: llm/leaderboard.json, llm/models/<id>.json. MCP tool: llm_leaderboard(sort?). Probing skips cleanly when no keys are set; spend is capped at $2.50/30 days in llm/models.json.