Ask your cluster (MCP)

WatchLog has a built-in MCP server. Connect your AI assistant to it, and you can ask questions like:

• "What's wrong with my clusters?"
• "Why does the cart pod keep restarting?"
• "Which containers on docker-01 were killed for memory today?"
• "Did anything change right before the errors started?"

The assistant answers from what WatchLog sees: problems and alerts, pods and containers, exit codes and OOM kills, events, recent rollouts, the service map and metrics. It works for Kubernetes clusters and Docker hosts alike.

WatchLog doesn't run an AI model itself. It only answers the questions your assistant asks through MCP. Nothing about your clusters leaves WatchLog unless you connect a client.

Try it on the live demo

The live demo has a public, read-only token. In a terminal with Claude Code:

claude mcp add --transport http watchlog-demo https://demo.watchlog.io/mcp \
  --header "Authorization: Bearer wl_pat_demoWafe1IZZ5LgmkIMrbhhfs932B8pd75Lr8P4fomh"

Then ask "What's wrong with this cluster?". The demo token reads only the sample cluster. It allows 30 tool calls a minute per visitor and 600 a minute in total.

Connect your WatchLog

You need:

• A personal API token. In WatchLog, open Settings → API tokens and choose Create token. Use one token per app, so you can revoke one without breaking the others.
• WatchLog's address, reachable from the machine your assistant runs on. That's the address in Settings → General → Public address. If WatchLog is only on a private network, connect through a VPN or a tunnel first.

The MCP server is at <your address>/mcp, for example https://watchlog.example.com/mcp. After you create a token, WatchLog shows the setup for each app below with your address and token filled in.

Claude Code

claude mcp add --transport http watchlog https://watchlog.example.com/mcp \
  --header "Authorization: Bearer wl_pat_..."

Claude Desktop

Claude Desktop connects through mcp-remote, which needs Node.js. Open Settings → Developer → Edit config and add:

{
  "mcpServers": {
    "watchlog": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://watchlog.example.com/mcp",
        "--header",
        "Authorization: Bearer ${WATCHLOG_TOKEN}"
      ],
      "env": { "WATCHLOG_TOKEN": "wl_pat_..." }
    }
  }
}

Restart Claude Desktop. The token stays in the environment variable, out of the command line.

ChatGPT

The ChatGPT desktop app and the Codex CLI read their MCP servers from ~/.codex/config.toml. Add:

[mcp_servers.watchlog]
url = "https://watchlog.example.com/mcp"
bearer_token_env_var = "WATCHLOG_TOKEN"

Then set WATCHLOG_TOKEN to your token in the environment the app starts from. If the app doesn't see the variable, put the header in the file instead:

[mcp_servers.watchlog]
url = "https://watchlog.example.com/mcp"
http_headers = { "Authorization" = "Bearer wl_pat_..." }

Restart the app, and WatchLog shows up under Settings → MCP servers.

ChatGPT on the web adds MCP apps in developer mode. Those apps sign in with OAuth, which WatchLog doesn't offer yet, so use the desktop app or Codex for now.

From your own code, OpenAI's Responses API takes WatchLog as a remote MCP server: {"type": "mcp", "server_label": "watchlog", "server_url": "https://watchlog.example.com/mcp", "authorization": "wl_pat_..."}.

Cursor

Add to ~/.cursor/mcp.json, or to .cursor/mcp.json in a project:

{
  "mcpServers": {
    "watchlog": {
      "url": "https://watchlog.example.com/mcp",
      "headers": { "Authorization": "Bearer wl_pat_..." }
    }
  }
}

Other clients

Any MCP client that supports Streamable HTTP and a custom header works. Send Authorization: Bearer <token>. WatchLog speaks the 2026-07-28 protocol, and the 2025 versions (2025-11-25, 2025-06-18, 2025-03-26) for clients that haven't moved yet.

What the assistant can do

Every tool is read-only. Each takes the cluster by name, and answers in that cluster's words: namespaces, workloads and pods on Kubernetes; hosts, Compose projects and containers on Docker.

Two prompts start an investigation for you: investigate_problem and why_is_it_crashing.

Limits:

• Tool calls: 60 a minute per token.
• Answer size: at most 16 KB each, with a fixed number of items. When an answer is cut, it says so, and which argument narrows it (for example "pass namespace").
• Metrics: at most 24 hours, 10 series and 60 points per series.

Security

• Read only. A token can't change anything in WatchLog or in your clusters, and no tool writes. The worst a misled assistant can do with WatchLog is read what your token reads.
• What it never returns: Secrets, environment variable values, request or response payloads, alert channel targets, email or Telegram settings, and tokens. WatchLog never collects most of these in the first place.
• Cluster text is data. Pod and container names, event messages, labels and images come from whoever runs workloads in your cluster. Someone could name a pod "ignore previous instructions…" to try to steer your assistant (prompt injection).
  • WatchLog returns such text only in its own fields, cleaned of control characters and cut to a fixed length.
  • Every tool tells the assistant to treat it as data.
  • Since the tools only read, an injected instruction can't make WatchLog do anything.
  • It could still try to steer the assistant's other tools in the same session. Review what your assistant does with those.
• Your call log. Settings → API tokens → a token shows its last 50 tool calls: the tool, a short summary of the arguments, the status, the time and the size, never the answer. Calls are kept 30 days.
• Revoking a token stops it at the next call.

Turn it off

The MCP server is on by default, since it needs a token anyway. To turn it off, set mcp.enabled: false in the Helm chart, or WATCHLOG_MCP_ENABLED=false with Docker Compose. /mcp then answers 404. Tokens still work with the REST API.

Troubleshooting