All selected work
05 / Case study Tool access · Personal system

One gateway for every AI assistant’s tools.

I put every connection between my AI assistants and outside systems behind one local gateway, Toolport. I didn’t build it. I chose it, configured it, and decided which of its protections were worth the friction.

Where it stands One place to grant and revoke · scoping stronger on paper than in practice
System / 05Sole operator: chose the gateway, configured connections, access profiles and security settings
01 Problem → decisions → result
The problem

What needed
solving.

Every assistant wants its own copy of every connection (a database, a notes vault, a ticket tracker, a design tool, store back ends), and every project folder is another place to decide access. Take an agent off a project and the access doesn’t leave with it, because it was attached to the chat window, not the project. And every tool description costs context whether or not it gets used.

The calls that mattered

What I decided,
and why.

  1. 01

    One gateway instead of per-assistant installs, so granting and revoking access is one edit in one place, and switching providers stops being a migration.

  2. 02

    Search for a tool instead of loading them all. Assistants get three small tools (search, call, status) instead of the full catalog at the start of every session. The cost: a search that comes back empty reads like the capability doesn’t exist.

  3. 03

    Access follows the folder, not the chat. Profiles are bound to working directories. That is the part I care about most.

  4. 04

    Keep the protections that cost nothing at the moment of use. I left off the ones that would put a prompt in front of me dozens of times a day. That is the weakest part of the setup, and the first thing I’d change before putting anyone else’s data behind it.

Where it landed

What shipped,
and what isn’t proven.

Every assistant I use talks to the gateway, it survives a change of provider without a reinstall, and unused tool descriptions stay out of context. Not proven: I have never pulled access from a running agent to watch it fail closed, per-assistant scoping is configured but empty, and most of my project folders still point at the general profile, so the scoping argument is stronger on paper than in my own setup.

What I’d change

What I’d do
differently.

Bind profiles to folders from the start, instead of leaving most projects on the general profile. And check the registry more often: one server is registered twice under two different names.

Stack
ToolportMCPClaudeAntigravity
Next case study A roadmap tool deliberately not synced to anything.