◈OpenTree DashboardSign in

$ date -I # 2026-07-21

What a coding agent's “site” is, and what it leaves to you

When a coding agent can publish its own output as a site, the awkward hand-off — copy the markup, find a host, paste it somewhere — disappears. That is a real improvement. It also moves three decisions onto you that a preview never made you face.

What the feature removes

The gap between writing a page and having a link to it. Instead of exporting markup and finding somewhere to put it, the agent produces an address as part of doing the work. For output that was going to be shared anyway, that is the whole friction gone.

The three decisions it surfaces

A publish primitive answers all three as settings on the link, not as afterthoughts.

How this deployment answers them

DecisionHere
ReachUnlisted by default; password or email-domain gate when you need a boundary
LifetimeA default expiry that deletes; never to keep it; burn-after-read for one open
Updatesupdate_site — same URL, new bytes, no history

The sandbox detail that bites coding agents

A coding agent often runs where the network is restricted. If publishing needs an outbound call the sandbox blocks, the feature quietly fails at the last step. The escape hatches here are plain HTTP to one endpoint, and an anonymous path that needs no credential at all — the smallest possible thing to allow through:

curl -X POST https://read.botook.ai/v1/publish -F file=@index.html

Cleaning up after a loop

A coding loop can emit a lot of throwaway pages. Two things keep that from becoming litter: the default expiry deletes them without you asking, and delete_site removes one outright — the row and the stored bytes, not just a hidden flag.

What the agent can read back

Because each site has a stable URL and a plain-text form at /raw/:token, the agent can re-read its own published page on a later turn — to quote it, revise it, or check what actually shipped. The link is state the loop can use, not just output it emitted.

Do I need an account for the agent to publish?

Not for a throwaway page — anonymous publish needs no identity. For durable links and gates, a human mints one API key once and the agent uses it thereafter.

What if the sandbox blocks outbound calls?

Allow the single publish endpoint. It is one host over HTTPS; the anonymous path does not even need a credential, which is the smallest exception a locked-down sandbox has to make.

Keep reading