A few weeks ago I wrote here about safe-by-default MCP servers for infrastructure and the governance model behind them. Since then the suite has grown to 15 servers — and the interesting part isn't the count, it's what happens when you apply one governance model to very different surfaces: clusters, databases, mail, chat, files, your own documents, and an entire Mac.
Here's what's new, and what building it taught me.
The governance model (quick recap)
Every server shares one small, pure, unit-tested policy engine:
- Read-only by default — write/admin tools aren't even registered until you raise the access mode, so the model can't call what it can't see.
- Allowlists + protected resources — scope to specific namespaces / folders / teams; mark things (kube-system, a prod DB, Sent Items) readable-but-never-mutable.
- Delete gating — destructive ops need an explicit opt-in flag on top of admin mode.
- Human-in-the-loop — destructive and high-impact actions pause and ask a person to approve, live, via MCP elicitation. A model can echo a confirmation string; it can't click a button that's shown to a human.
- Dry-run, JSON audit to stderr, secret redaction on top.
New: browser sign-in for Microsoft 365
The newest servers — mcp-outlook (mail and OneDrive files) and mcp-teams — talk to Microsoft Graph. The hard part for a local stdio server is auth: you don't want to paste bearer tokens or handle a password.
So they do an az login-style flow. On first run the server opens your browser to the Microsoft sign-in page (OAuth authorization-code + PKCE with a one-shot loopback redirect on 127.0.0.1), catches the redirect, exchanges the code, and caches the access + refresh token on disk at 0600. After that it refreshes silently — the browser only reappears if the refresh token is gone. The server never sees your password.
The safe-by-default model still applies on top: read-only until you set OUTLOOK_MODE=read-write; sending mail needs an extra OUTLOOK_ALLOW_SEND flag and a human confirmation, because a sent message can't be recalled; deletes go to the recycle bin, gated behind admin + OUTLOOK_ALLOW_DELETE.
New: ask your own documents (RAG) + agent memory
ossian-mcp lets an agent query your own documents and get answers with the exact passages they came from, plus durable memory ranked by relevance, importance and recency. Same flag family — OSSIAN_ALLOW_FORGET gates the one destructive tool.
New: controlling a Mac like a human
Two macOS servers, and they're a nice illustration of the model. mcp-macos is safe-by-default: observe the machine, and only behind explicit opt-ins run shell / AppleScript / GUI input. mcp-mac-control is the deliberate opposite — full control out of the box, with a single MACCTL_SAFE_MODE=true switch that inverts it back to locked-down.
The full-control one drives the Mac like a person: screenshot to see the screen, then move, click (left/right/double), drag, scroll, type, and key combos — plus shell and AppleScript for the things a mouse can't reach. The lesson here was worth the whole exercise: keep the same pure policy engine, and flip only the defaults. The identical code path runs wide open or fully gated depending on one flag.
What I'd tell my past self
- The governance model transferred cleanly from Kubernetes to a mailbox to a whole desktop. The abstraction that mattered was "capability + access mode + per-op flag," not anything domain-specific.
- Human-in-the-loop via elicitation is the highest-leverage safety feature — it's the one thing a model genuinely can't fake.
- Browser sign-in beats token-pasting for anything a human owns.
The whole suite — 15 servers, MIT/Apache, on npm and the official MCP registry — is on one page: https://dockndevai.github.io. Repos: https://github.com/dockndevai.
I'd love feedback, especially on the last one: for desktop control, is full control by default with an opt-out or safe by default with opt-ins the right posture? I shipped both and I'm genuinely not sure which becomes the default people reach for.