MCP FOR YOUR AGENT

Your agent reads production instead of guessing.

Connect Kopai once. Claude, Cursor, Codex or any other MCP tool answers from your app's traces, logs and metrics.

  • Read-only on your data
  • No API key to copy
  • Set up in two minutes

An illustrated Claude chat with Kopai connected. It plays example chats one at a time: an incident triaged from error rates and logs, a slow route found from today's traces, and one request followed through its logs. Pick a chat to read it at your own pace.

Checkout errors

Checkout is failing for some users since this morning. What's going on?

querytraces · error rate by service · last 3h
querylogs · checkout-api · ERROR · last 3h

From the last three hours in Kopai:

  • Failing: 4% of POST /api/checkout return 504, all since 2:14 AM.
  • Cause: every failed request times out in stripe.charges.create after 30 s. There is no timeout guard on that call.
  • Who: 23 users so far, all paying with the new pay-later option.
  • Elsewhere: payments, auth and search are clean.

Want me to draft the fix?

KopaiReply to Claude…

Sounds familiar?

  • Logs pasted into every chat
  • Screenshots of dashboards
  • “Can you paste the error again?”
  • Redeploy and hope
  • Guessing from the code

Every chat starts blind.

Your AI tool never saw the request, the trace or the error.

WHAT YOU CAN ASK

Ask about production, in plain language.

Your AI tool turns the question into Kopai queries and follows up on what it finds.

Triage an incident

“Which services started returning errors in the last 30 minutes, and what are the most common messages?” It compares error rates by service, then reads the matching traces and logs.

checkout-api4.1%
payments0.2%
search0.1%
auth0.0%
top messagestripe.charges.create timed out after 30s· 23

Knows your services

Every service, route and metric name, discovered from what you already send. Nobody types them into a prompt.

DISCOVERED

7 services, 48 routes, 31 metrics. http.server.duration by route and status. checkout.orders.total by plan.

Follow one request

Logs carry the trace ID of the request that wrote them, so it lines them up with the spans, service by service.

TRACE 4BF92F35…

api-gateway200 msrequest received, routed to orders
orders1.8 sinventory lookup retry 3/3 · fallback to cached stock
inventoryfailedconnection refused 10.0.4.12:5432

Debug next to the code

In Claude Code, ask “why is this handler slow in production?” next to the handler. It reads the handler's traces and points at the call that takes the time.

POST /signup2.9 s
hashPassword120 ms
db.insert users48 ms
sendWelcomeEmail2.3 s

A report every morning

A scheduled task in Claude checks error rate and p95 by service and posts a short summary, even while your laptop is closed.

SCHEDULED, 9:00 AM

Last 24 h: checkout-api error rate 0.3%, down from 4.1%. p95 420 ms. No new error messages. Nothing to do.

A chart that stays current

Ask Claude Code to “keep an eye on error rate by service”. It publishes a live page that re-queries Kopai each time someone opens it, so the team looks at the same numbers without re-asking.

Error rate by service· last 24 hLIVE

Refreshed 30 s ago · each viewer's own connector, never a copy of the data

THE FAIR QUESTION

Claude already reads my code.
Isn't that enough?

Code says what should happen. Kopai says what did.

You ask once. It checks production itself.

Without Kopai

Checkout is failing for some users. Here is this morning's log export [CSV pasted], the Stripe dashboard [screenshot] and the last deploy diff [pasted]. Context: Next.js on Vercel, Stripe for payments, pay-later launched on Friday, target p95 is 800 ms. Can you work out what is wrong and which users it hits? And remind me to check again tomorrow, I will forget…

Opus ▾

Export, paste, explain the app again.

Claude, with Kopai connected

Why is checkout failing for some users?

querytraces · error rate by service · last 3h
querylogs · checkout-api · ERROR · last 3h

4% of POST /api/checkout return 504 since 2:14 AM, every one timing out in stripe.charges.create after 30 s. 23 users so far, all on pay-later. Want the fix?

It runs the queries itself.

See everything you can ask →
PICK YOUR TOOL

Setup steps for your tool.

One address. Add it in your tool and sign in.

https://api.kopai.app/v2/mcp

Connect in Claude.ai and Claude Desktop

Kopai is in the Claude connector directory. Connect it there once; Claude Desktop uses the same connector as your claude.ai account.

  1. 1.

    On the listing, choose Connect to Claude.

  2. 2.

    Sign in to Kopai if asked, then choose Approve on the consent page. On Team and Enterprise, an owner can add it once for everyone under Organization settings › Connectors.

Add Kopai from the directory only. A custom connector with the same address counts as a second Kopai, and live pages work with exactly one.

CLI tokens, the local app and troubleshooting are in the connector docs.

SETUP

Set up in two minutes.

No API key to create. Add the address, sign in, approve.

  1. 1

    Add the address in your tool

    Kopai is in the Claude connector directory, one click away. Claude Code, Cursor, Codex and any other MCP host take the same address.

    >_
    api.kopai.app/v2/mcp
  2. 2

    Sign in and approve once

    Your tool sends you to the Kopai portal. The consent page names the app, the workspace and where you'll be sent back.

    CONNECT AN APP

    Claude is asking to read the telemetry and alerts of acme-prod.

    ApproveDeny
  3. 3

    Ask your first question

    In plain language. Your tool picks the Kopai tool and writes the query; you read the answer.

    what's erroring right now?

    checkout-api: 4.1% of requests since 2:14 AM, all timing out in Stripe. Everything else is clean.

Add to Claude
YOUR DATA

One workspace. Your data is never changed.

It reads your telemetry and alerts. The only thing it can change is alert rules, and only for owners and admins.

workspace it can reach
1
API keys to copy
0
until a revoked app stops
30 s

Where your data goes

  1. Stored in the EU

    Your telemetry lives at Hetzner in Germany and Finland. The connector doesn't move it.

  2. Queried in the EU

    When someone asks, the query runs where the data is.

  3. Only the result leaves

    At most a few hundred rows go to your AI host, which shows them to you.

For Claude, that host backend is Anthropic's, in the US. The host's own data policies apply to what it receives.

  • Reads your data, changes only alert rules

    It reads the telemetry and alerts of your workspace. Owners and admins can also create and change alert rules. It cannot change data, members, billing or settings.

  • No key to copy

    Sign in with your Kopai login and approve the app once. Its access refreshes on its own until you revoke it.

  • Nothing sent unless asked

    A query runs when someone asks a question, and only its result reaches the AI host.

  • Revoke anytime

    Account › Connected apps in the portal lists every app you've approved. Revoke one and it stops within 30 seconds.

PRICING

Included in every plan.

No AI add-on. No per-seat price.

FREE

$0/mo

for the whole team

Unlimited seats, unlimited questions, every signal, free today.

Included

  • The connector in every plan, no AI add-on
  • Every seat included, never per-seat
  • Unlimited API calls and questions
  • Storage-based pricing when paid plans arrive
  • The same MCP server in the free local app

What a question costs

  • An incident triage

    $0

    included

  • A slow-route hunt

    $0

    included

  • A trace walk-through

    $0

    included

  • A morning report

    $0

    included

See full pricing →

Questions, answered.

Which AI tools work?

Claude on claude.ai and Claude Desktop, Claude Code, Cursor and Codex have setup steps on this page. Any other MCP host that supports remote servers with OAuth sign-in takes the same address. For Claude, Kopai is in the connector directory.

Do I need an API key?

No. You sign in with your Kopai login and approve the app once; its access refreshes on its own until you revoke it. Claude Code on a machine without a browser can use a CLI token instead.

What does my AI tool see?

The traces, logs, metrics and alerts of one workspace, as the result of a query, at most a few hundred rows, and only when someone asks. Your data stays in the EU; the result goes to the AI host, whose own data policies apply to it.

Can it change anything?

Only alert rules, and only for workspace owners and admins. It cannot change telemetry, members, billing or settings, and it cannot reach any workspace but the one you approved.

How do I turn it off?

Account › Connected apps in the Kopai portal lists every app you've approved. Revoke one and it stops working within 30 seconds.

What does it cost?

Nothing on top of Kopai Cloud, and Kopai Cloud is free today, no credit card. When paid plans arrive, one simple price that grows with storage, never per seat.

Does it work with the local Kopai app?

Yes. The local app serves the same server at localhost:8000/mcp for the telemetry on your machine, with no sign-in, for Claude Code and other clients on that machine. Details are in the connector docs.

FREE TO START

Ask your AI tool about production.

Connect Kopai once. It works in Claude, Cursor, Codex and any other MCP tool.

No credit card  ·  Read-only on your data  ·  Revoke anytime