Your AI editor can see your production

Install the Safeship MCP server and Claude Code stops guessing. It creates your monitors, reads your incidents and tells you what actually broke.

you

the signup endpoint feels broken since this morning, what happened

claude

Calling get_incident_report. Your POST /api/signup check started failing at 09:12. It still returns 200, but the response body lost the userId field, so the schema check fails. Twenty-three consecutive runs affected. The last deploy went out at 09:07.

Install it in one command

The server is hosted, so there is nothing to install and nothing to keep up to date. Generate an API key under Settings, then API Keys, and point your editor at it. In Claude Code:

claude mcp add --transport http safeship \
  https://mcp.safeship.app/mcp \
  --header "X-API-Key: sk_live_your_api_key"

Any other MCP client takes the same thing as a config entry:

{
  "mcpServers": {
    "safeship": {
      "type": "http",
      "url": "https://mcp.safeship.app/mcp",
      "headers": {
        "X-API-Key": "sk_live_your_api_key"
      }
    }
  }
}

What your editor can do once it is connected

create_check

Create a monitor on an endpoint, with the thresholds it should hold to.

discover_and_create_check

Point it at an API and let it work out what to monitor on its own.

analyze_endpoint

Inspect an endpoint and infer what a valid response looks like.

get_incident_report

Pull what broke, when, and what the endpoint returned at the time.

get_health_score

Read the current health of everything you monitor, in one number.

get_analytics

Success rate, latency and error trends over a period.

list_alerts

List the alerts that fired, so your editor can explain the last incident.

create_channel

Add an alert destination: email, webhook, SMS, WhatsApp.

Why this is different from another dashboard

Monitoring fails for solo developers for a boring reason: setting it up is a chore you do once, badly, and then never revisit. Every endpoint you add afterwards goes unmonitored because opening the tool, filling the form and picking the thresholds is friction you will not pay again.

When the setup happens where you already are, that friction disappears. You ship an endpoint, you say monitor this, and it is monitored. When it breaks three weeks later, you ask what happened in the same window where you are about to write the fix.

Not a monitor for MCP servers

Worth being explicit, because the terms are close. Safeship watches your production APIs and websites. The MCP server is the door your AI editor walks through to reach that monitoring. It is the interface, not the subject.

What it checks, not just whether it answers

Safeship validates the response body against a schema, so an endpoint returning 200 with missing or malformed data counts as a failure. That is the class of bug that survives every ping-based monitor, and the one your AI editor is best placed to explain once it can read the actual payload.

Common questions

What is the Safeship MCP server?
An MCP server that exposes your Safeship monitoring to any MCP-compatible AI editor. Your assistant can create checks, read incident reports and pull your health score without you opening a dashboard.
Which editors does it work with?
Any MCP client. Claude Code, Claude Desktop and Cursor are the ones people use most. The server is hosted and speaks streamable HTTP, so installation is one command or one config entry, with nothing to install locally.
How do I authenticate the Safeship MCP server?
With an API key sent as the X-API-Key header. Create one under Settings, then API Keys, at safeship.app. Keys start with sk_live_. Nothing else is needed, there is no OAuth flow to complete and no token to refresh.
Does it monitor MCP servers?
No, and the distinction matters. Safeship monitors your production APIs and websites. The MCP server is how your AI editor reaches that monitoring, not the thing being monitored.
Can my AI editor see my production errors?
Yes. That is the point. Ask it why something broke and it calls get_incident_report, reads the failed responses and answers with the actual payload rather than a guess.
Is the MCP server included in the free plan?
MCP access is part of the paid tiers, which start at EUR 3 a month. The free tier covers 5,000 checks a month through the web app.

Let your editor set it up

Create an account, generate an API key, paste the config. Two minutes.

Get startedSee pricing