Introducing Familiar: the in-app AI companion architecture
Familiar is the working name for an in-app AI companion architecture other apps can plug into — here's how it's designed, and where it stands today.
By Agent Software
An AI companion as a pluggable piece, not a bolt-on
Most in-app AI assistants get built once, specific to one product. Familiar takes a different approach: a companion architecture designed to be hosted by different apps, connected through a typed, versioned manifest rather than custom integration work for each host.
Internally, this is being built to support Agent Studio first, with the goal of the same underlying companion working across other apps in the suite over time.
Loopback-only, on purpose
The companion service runs on 127.0.0.1 — loopback only, no external network exposure by default. A host app authenticates with it locally, which keeps the trust boundary narrow: only the app actually running on the same machine, at the same time, in the foreground, can talk to it.
What it can actually do
Familiar supports a few concrete interaction types: highlighting relevant UI in the host app without simulating clicks or pointer input, explaining what's on screen using signed, checksummed documentation bundles rather than free-form guesses, and proposing multi-step actions through a plan-approve-execute flow.
That last part matters most. Nothing gets executed without the user approving a plan first, and approvals are time-boxed rather than standing permissions — a five-minute lease, not a blank check.
Safety by design, not by policy
A few of the guardrails worth naming: high-risk commands require double confirmation, not a single click-through. Every proposed action needs a defined rollback strategy before it's allowed to run. Activity history stores only metadata about what happened, never the actual prompts or parameters involved. And the companion only responds to the foreground host app, closing off a class of cross-app privilege issues.
Where this stands today
Familiar is functionally built — the engine, the approval flow, the safety checks, and an integration with Agent Studio are in place and have been through internal testing. It is not a publicly available product yet: there's no public app, no signup, and remaining release-readiness work (a full regression pass, accessibility review, final sign-off) isn't done. This post describes the architecture as it's being built, not something available to try today.
