okokumo in your AI assistant

okokumo runs an official, hosted Model Context Protocol server. Connect it to an AI assistant and you can ask which checks are down, whether you were alerted, what a check's uptime was this week — or have it create a check, pause one for a deploy, or narrate an incident on your status page.

It is a thin layer over the public REST API: every tool is a call to a /v1 route, made as you — through a browser sign-in or an API key — with the same authentication, organization scoping, role checks and plan limits. It can do nothing an API key cannot already do over REST.

• Endpoint: https://app.okokumo.com/mcp, MCP Streamable HTTP
• Authentication: browser sign-in (OAuth), where the client needs only the URL; or an okokumo API key, sent as Authorization: Bearer sk_…
• Plans: all of them, Free included
• Rate limit: your organization's API budget, 120 requests per minute, shared with REST

1. Connect by signing in (recommended)

Claude, ChatGPT, Cursor, VS Code and Claude Code can connect with nothing but the URL https://app.okokumo.com/mcp. There is no key to create, paste or keep out of git.

1. Add okokumo to your client with that URL (per client below).
2. The client opens okokumo in your browser. Sign in if you are not already.
3. A consent screen shows the app's name, the address it will send you back to, your organization, and what the app gets: full access, the same as an API key, with your role. Approve or deny. Only approve a connection you started yourself.
4. Back in the client, the tools are there.

Only owners and admins can approve, the same rule as for creating an API key. The app is connected to the organization named on the consent screen, and to that one only.

Claude (claude.ai, Claude Desktop)

Settings → Connectors → Add custom connector, with the URL https://app.okokumo.com/mcp.

ChatGPT

In ChatGPT's connector settings, add a custom connector with the same URL.

Claude Code

claude mcp add --transport http okokumo https://app.okokumo.com/mcp

Then run /mcp in Claude Code, pick okokumo and sign in.

Cursor

In ~/.cursor/mcp.json or .cursor/mcp.json:

{
  "mcpServers": {
    "okokumo": { "url": "https://app.okokumo.com/mcp" }
  }
}

Cursor then offers to sign in from its MCP settings.

VS Code

In .vscode/mcp.json:

{
  "servers": {
    "okokumo": { "type": "http", "url": "https://app.okokumo.com/mcp" }
  }
}

VS Code asks you to sign in when it starts the server.

Tested end to end so far: Claude Code and the MCP Inspector. claude.ai, ChatGPT, Cursor and VS Code implement the same MCP authorization specification but have not yet been tried against okokumo live. If one of them fails, tell us; an API key (section 3) works in the meantime.

2. Connected apps and revoking

Account → Connected apps lists every app connected to your organization: its name, who connected it and when, and when it was last used. Every member can see the list; owners and admins can Revoke an app there. Revoking is immediate: the app is refused at its next request.

• Access tokens last an hour and renew on their own. An app that goes unused for 30 days has to be approved again.
• A password reset, a suspended account or removal from the organization disconnects every app that person connected.

3. Or use an API key

For scripts, CI, and clients without OAuth sign-in. API keys work exactly as before, alongside connected apps.

In the dashboard, Account → API keys (owners and admins). The key is shown once.

• A key acts with the role of the person who created it. An owner's or admin's key can do everything listed below; a member's key can manage checks but not incidents — exactly as over REST.
• Use one key per client — one for Claude Code, one for Cursor — so each can be revoked on its own. The key is checked on every request, so a revoked key cuts the client off at its next message.
• The configuration below holds a credential. Treat it like a password file and keep it out of git.

4. Connect a client with a key

Replace sk_… with your key.

Claude Code

claude mcp add --transport http okokumo https://app.okokumo.com/mcp \
  --header "Authorization: Bearer sk_…"
claude mcp get okokumo

For a project-wide .mcp.json (--scope project), put Bearer ${OKOKUMO_API_KEY} in the header and export the variable in your shell, so the key never lands in the repository.

Claude Desktop

The custom connector in section 1 needs no key. To use a key instead, connect through the mcp-remote bridge (needs Node.js) in claude_desktop_config.json:

{
  "mcpServers": {
    "okokumo": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://app.okokumo.com/mcp",
               "--header", "Authorization:${OKOKUMO_AUTH}"],
      "env": { "OKOKUMO_AUTH": "Bearer sk_…" }
    }
  }
}

The header value goes through env because some platforms split arguments on the space in Bearer sk_…. Restart Claude Desktop after editing.

Cursor

In ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project):

{
  "mcpServers": {
    "okokumo": {
      "url": "https://app.okokumo.com/mcp",
      "headers": { "Authorization": "Bearer sk_…" }
    }
  }
}

Other clients

Any MCP client that speaks Streamable HTTP and can send an Authorization header connects the same way: the endpoint above, your key as a bearer token.

5. What it can do

18 tools: 10 that read, 8 that write. Nothing deletes, and nothing touches API keys — deleting anything and managing keys stay in the dashboard.

• Read: your account; your checks; a check's uptime and latency over a day, week or month; its state changes; the alerts it sent; the channels it alerts; your channels; your status page and its incidents.
• Write: create a check, update it, pause it, resume it, choose which channels it alerts; create an incident, update it, post an update to it.

What differs from calling REST yourself, on purpose:

• Partial updates. Editing a check or an incident changes only the fields you name; the tool reads the current record first and keeps the rest.
• Durations are seconds. grace_seconds: 600 is ten minutes. The tool descriptions say so, because minutes are the classic model mistake here.
• Ping URLs included. Heartbeat and cron checks come back with the ping_url the job has to call.
• Channel secrets stay out of the transcript. Webhook URLs, signing secrets and Telegram chat ids are never returned.
• Strict arguments. An unknown argument is refused with the list of accepted ones rather than silently dropped.
• Writes are marked. Every tool declares whether it only reads or what it changes, so your client can ask before running a write. Pausing a check or narrowing its alert routing counts as destructive — it means alerts that will not be sent — and the model is told to confirm with you first, and to publish an incident only when you ask: incidents start as private drafts.

6. Limits and errors

• Every request to /mcp counts against your organization's 120 requests per minute — the same budget as REST, and the two add up. Tools that read before they write (updating, pausing or resuming a check, updating an incident) cost two.
• Over the limit, the assistant is told how many seconds to wait, and not to retry in a loop.
• A refused call — a plan limit, a role that is not allowed, an id that does not exist — comes back as an error the model can read and explain to you, ending with the HTTP status.
• An unknown or revoked key gets 401: create a new key and update the client's configuration. A revoked app gets the same answer and has to be approved again.