AIOps Field Notes

AIOps Field Notes MCP server

Four things I kept doing by hand while running AI in production, turned into tools an agent can call: price a model before committing a budget, size a GPU before ordering one, review an MCP config before trusting it, and explain what changed in a network device config.

Free, MIT, no account, no telemetry.

Install

Pin the version — see audit_mcp_config for why.

Claude Code

claude mcp add aiops-field-notes -- npx -y github:labaccessnow/aiops-field-notes-mcp#v0.1.1

Claude Desktop, Cursor, or any client with a JSON config

{
  "mcpServers": {
    "aiops-field-notes": {
      "command": "npx",
      "args": ["-y", "github:labaccessnow/aiops-field-notes-mcp#v0.1.1"]
    }
  }
}

Docker

{
  "mcpServers": {
    "aiops-field-notes": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "ghcr.io/labaccessnow/aiops-field-notes-mcp:0.1.1"]
    }
  }
}

Node 18 or newer for the npx route; nothing else to configure either way. The server is also in the official MCP registry as io.github.labaccessnow/aiops-field-notes-mcp.

Tools

audit_mcp_config

The one I use most. Paste a client config and it reads it the way a reviewer would — every server entry is an access grant, so the questions are what it can reach, who publishes it, and what happens the next time it auto-updates:

6 servers reviewed — 7 high, 6 medium, 0 low.

files
  [high] Package version is not pinned
    files runs "@modelcontextprotocol/server-filesystem" at whatever version is published when it launches.
    fix: Pin it — "@modelcontextprotocol/server-filesystem@<version>" for npm, "@modelcontextprotocol/server-filesystem==<version>" for uvx — and bump deliberately.
  [high] Filesystem root is wide open
    files is granted "/" — that is the whole home directory or the whole disk, including keys and browser profiles.
    fix: Scope the server to the project directories the agent actually works in.
  [medium] Installs without prompting
    files passes -y, so a new release installs and runs with no confirmation.
    fix: Keep the flag only alongside a pinned version, so what installs is what you reviewed.

notes
  [high] Package version is not pinned
    notes runs "some-random-notes-mcp" at whatever version is published when it launches.
    fix: Pin it — "some-random-notes-mcp@<version>" for npm, "some-random-notes-mcp==<version>" for uvx — and bump deliberately.
  [high] Secret is stored in the config file
    notes has NOTES_API_KEY written out in plaintext; config files get committed, synced and screenshared.
    fix: Reference the environment instead — "NOTES_API_KEY": "${NOTES_API_KEY}" — and keep the value in your secret store.
  [medium] Installs without prompting
    notes passes -y, so a new release installs and runs with no confirmation.
    fix: Keep the flag only alongside a pinned version, so what installs is what you reviewed.
  [medium] Publisher not in the tracked set
    "some-random-notes-mcp" is not one of the reference or vendor servers this directory tracks.
    fix: Read the source and check the publisher before wiring it into an agent that touches production.

It also flags shell entrypoints, privileged containers, host-root mounts, credentials on the command line, database access that is not read-only, unauthenticated remote servers, and anything reached over plain HTTP.

What it does not do

• No account, no signup, no key.
• No telemetry. Nothing about your usage leaves your machine.
• No filesystem access. audit_mcp_config takes the config as text, so the server never reads your disk.
• One network call in the whole package: latest_field_note fetches a public RSS feed. Skip that tool and this server never opens a socket.

The cost table, GPU list and server directory match the calculators at aiopsfieldnotes.com/tools and get re-verified with each weekly episode. Pricing moves — check the vendor's page before you commit a budget to it.

Where the network tools come from

explain_config_diff and check_config_compliance are the read-only core of DriftWatch, which runs the same rules nightly across a fleet and keeps the history. Here you get the single-shot version, on two snapshots you paste, with no storage and no scheduler. That is deliberate: the rules are the useful part, and they work fine on their own.

Licence

MIT. Written by James Son — network, security, and automation engineer — and tested in a live multi-vendor lab. Contributions and corrections welcome.