Proposal: a terminal ability for Littlebird — sanctioned command execution on the user’s desktop

For: littlebird.ai product/engineering team From: Christopher Wilson, Founding Brand Ambassador Date: October 27, 2026 Status: Draft — for Christopher to review before sending

Summary

Littlebird watches my desktop — screen context, files, meetings, messages — but she cannot act in the one place most technical work actually happens: the terminal. Today every verification loop is a relay race. I run the command, screenshot or copy the output, paste it back into chat, and she reasons over a photograph of the truth. I’d like to propose a sanctioned terminal ability: Littlebird executes commands on my machine, through an explicit approval flow, with scoping and an audit trail.

The problem, concretely

While building the Taylor’d Wraps interactive preview with Littlebird over the past two days, I lost count of the relay-pastes:

“Run gh repo edit ... and tell me what it says” — I run it, paste the interactive prompt screenshot, she reads it, tells me to spacebar-select Description and Topics.
“Check whether the Pages deploy finished” — I curl the URL, paste the output, she confirms.
“What did that Actions run fail on?” — I open the browser, screenshot the job steps, paste.

Every one of those was a round trip where Littlebird already knew exactly what to run and what the output would mean. The bottleneck was never knowledge; it was hands.

The ask

A device-scoped terminal ability (sibling to the existing Filesystem device integration) with four properties:

Approval-gated execution. Littlebird proposes a command as an action card — command, working directory, why — and it runs only when I approve. No background daemons, no unattended shells.
Scoped permissions. An allowlist posture: read-only commands (ls, cat, git status, curl GET) can be pre-approved per-session; mutating commands (git push, gh repo edit, file writes) always require a fresh approval. Sudo and secrets-bearing commands are refused outright.
Full audit trail. Every executed command, its output, and which chat/card triggered it is captured in the session record — so the observer is also observable.
A natural invocation syntax. Something like !gh run list in chat, so I can hand her a command in passing, and she can propose the same syntax back when she wants me to run something she can’t.

Why this helps littlebird.ai too

This is the difference between a Desktop Oracle Observer and a Desktop Oracle Operator. Every user who works in a terminal — developers, admins, data folks — hits the same relay bottleneck on day one. An approval-gated terminal ability with visible scoping is also a trust story: Littlebird can say “here is the exact command, here is why, approve to run” instead of “go run this and come back.” It is the most-requested gap I see in her otherwise complete observation stack.

What I’m not asking for

Not asking for unattended execution, sudo, persistent shells, or access to secrets/keys. The screen-observation flow stays; this adds a governed pair of hands, not a free-running agent.

Read more, or follow along: drasticstatic.github.io/littlebird-ambassador-public-preview