Giving Agents Eyes Without Giving Up Your Archive

The most capable models on earth cannot see your photos. That is not a limitation of the models — it is a consequence of where your files live.

You can fix it in one of two ways. Upload the archive and hand over custody. Or expose a local tool that lets an agent ask questions of an index that never leaves the machine.

The second option is the only one that scales to a real library, and it is mostly an interface problem.

An agent needs verbs, not a drive

Handing an agent a folder path is a dead end. It cannot meaningfully “list the directory” its way to the shot you remember. What it needs is the same primitives a person would use:

• Search by description, returning ranked candidates with enough context to judge.
• Look at a specific frame, at a resolution it can actually reason about.
• Read text that appears on screen, in any language it encounters.
• Hear what was said, aligned to time.
• Seek to a moment so the result is actionable.

Five verbs. Everything else is composition.

Consent has to be structural

The reason this is hard is not technical. It is that “let the assistant see my files” is currently a binary, permanent, and opaque grant.

Make it structural instead. Scope the tool to a folder, cap the resolution the agent receives, log every call locally, and let the grant be revoked without touching the index. The model gets to see the answer without holding the library.

Access should be a lease, not a transfer of title.

Why local wins here

A cloud index can answer these questions too — as long as you are willing to upload everything first, and stay inside the retention policy of whoever holds it.

A local index skips that negotiation entirely. The footage never moves, the permission lives on your machine, and the agent gets a window rather than a key. That is not a privacy feature bolted on at the end. It is the only version of this that stays true as the archive grows.

The archive is the asset. The agent should be a guest in it.