Semantic health monitoring for MCP

Catch the MCP failures your uptime monitor calls green.

mcpup runs the full initialize handshake, snapshots every tool's schema, and validates real responses — so silent failures, schema drift, and broken tools get caught before your agents do.

Start monitoring free $ Get the CLI

✓ Free for 3 monitors✓ No credit card✓ Open-source probe engine

Generic uptime monitor

payments-mcp 200 OK · Up

Looks healthy. It isn't.

payments-mcp

Silent failure

https://api.acme.io/mcp

Semantic checks · last probe

✓initializehandshake ok · protocol 2025-06-18

✓tools/list8 tools · schema unchanged

✕create_chargeJSON-RPC error inside a 200 OK

✓driftno schema change vs baseline

!Alerted Slack & webhook · 14:32 · agents paused

Built for teams self-hosting MCP forClaude Code·Claude Desktop·Cursor·custom agents

52%

of remote MCP servers in production are completely dead

are in full health — the rest are slow, stale, or silently failing

2,181

remote endpoints probed · April 2026

Source: April 2026 survey of 2,181 public remote MCP endpoints · read the methodology →

THE BLIND SPOT

Three failures Pingdom can't see

A generic monitor checks that a port answers. None of these trip it — every one breaks your agents.

Silent failures

A server returns 200 OK with a malformed payload or a JSON-RPC error inside. Uptime monitors see green. Your agent breaks mid-task.

200 OK → { "error": -32603 }

≠

Schema drift

A tool changes its inputSchema without warning. Agents that depend on it fail silently — often in prod, in front of a customer.

amount: string → integer

⋔

Sprawl & abandonment

Duplicate servers, expired upstream credentials, serverless cold starts that kill the first request. One platform lead found 14 servers — 4 unknown duplicates.

14 servers · 4 unknown dupes

WHY

We know what “healthy” means — semantically.

A generic pinger does an HTTP ping. mcpup speaks the protocol. That's the difference a commodity monitor can't copy — and it comes straight from hands-on agentic practice.

A generic pinger sees

200 OK

One number. No idea whether the handshake, the tools, or the responses actually work.

mcpup sees

Full initialize handshake

The complete MCP handshake over streamable HTTP — not just “the port answered”.

Schema snapshot & diff

Version every tool's schema via tools/list and diff each run to catch drift.

Safe tool call

Call a read-only tool with synthesized args and validate the shape of the response, not the HTTP code.

Real success ≠ 200 OK

Distinguish a genuine result from a 200 OK that hides a JSON-RPC error.

HOW IT WORKS

From URL to “I'll leave this running” in two minutes

Register your endpoint

Paste the MCP URL, add a bearer token or OAuth 2.1. mcpup probes it once and auto-detects the safest read-only tool.

https://api.acme.io/mcp

8 tools detected · list_charges selected

We probe it, semantically

Every 30s–5min: handshake, list tools, diff the schema, call the safe tool, parse the payload. Uptime, latency p50/p95, drift — all tracked.

p50 142msp95 410ms

Get told before agents break

Drift, silent failure, or downtime fires email, Slack, or a webhook — and feeds a public status page your customers can subscribe to.

#Slack · #mcp-alerts

@Email · on-call

⚡Webhook · pause agents

EVERYTHING IN THE BOX

A focused monitor — nothing you don't need

probe

Periodic semantic probe

initialize → tools/list → optional safe tool call, on your schedule.

drift

Schema drift detection

Structural diff against a versioned baseline. Alert the moment a tool's contract changes.

parse

Silent-failure detection

Strict JSON-RPC parsing surfaces errors and unreadable payloads hidden under a 200.

metrics

Uptime & latency

Uptime %, p50 / p95 latency, and an incident timeline per endpoint.

alerts

Multi-channel alerting

Email, Slack, and a generic webhook — wired to the events that matter.

status

Public status page

Prove your MCP's uptime to customers with a hosted, subscribable status page.

See a live example →

OPEN SOURCE · TOP OF FUNNEL

Start in your terminal.

The mcpup CLI is the open-source probe engine behind the cloud — a one-shot semantic health check with CI-ready exit codes. The hosted plans add scheduling, history, alerting, and dashboards.

✓ Silent-failure & drift detection, local

✓ Exit codes 0–4 for pipelines · --json output

✓ Built on the official MCP SDK

$bunx mcpup <url>

View on GitHub →

zsh — mcpup

$ bunx mcpup https://api.acme.io/mcp --call

→ initialize handshake ok · protocol 2025-06-18

→ tools/list ok · 8 tools · schema hash a3f9…7c2

→ drift none vs baseline

→ safe call list_charges …

✕ silent failure · 200 OK hiding JSON-RPC error -32603

outputSchema mismatch: expected object, got null

FAIL · 1 of 4 checks failed (exit 2)

PRICING

Priced per monitor. No sales call.

Bottom-up and self-serve. Start free, upgrade when drift history and a public status page earn their keep.

Free

$0forever

The magnet. Prove it on one server.

3 monitors

5-minute interval

1 alert channel

7-day history

Start free

Pricing questions

The stuff a skeptical dev wants answered before signing up.

What counts as a monitor?

One MCP endpoint URL. Each endpoint you watch is a monitor — no matter how many tools it exposes or how often it's probed.

What happens if I go over my plan's limit?

Your existing monitors keep running. You'll be prompted to upgrade before adding more — we never silently pause or drop a monitor you rely on.

Can I cancel anytime?

Yes. Monthly and annual plans are self-serve through Stripe — cancel in one click, no sales call, no retention maze.

Do you bill in USD or EUR?

Both — switch currency above the plans. Checkout is handled securely by Stripe in your selected currency.

Stop shipping agents that depend on broken MCP servers.

Join the waitlist and be first to add an endpoint — and know within minutes whether it's actually healthy, not just answering the door.