Initial problem
AI agents can produce useful plans, reports, prototypes, and dashboards as complete HTML files, but sharing those files is awkward. Local servers disappear, attachments lose their context, and generic static hosting does not provide private-by-default access, immutable history, or safe isolation for untrusted HTML. Agents needed a small publishing primitive designed around their actual output.
Implementation
AgentPlan accepts HTML from the browser, a scoped API, or a one-command CLI and publishes it behind a stable URL. Every draft keeps immutable version history and can be private, public, or password protected. Metadata and access policy live in PlanetScale Postgres while the HTML itself stays in a private Cloudflare R2 bucket. Content is streamed through a separately authorized route and rendered in a sandboxed iframe without same-origin access, so generated scripts never enter the application DOM.
Outcome
My contribution
Impact
What mattered
One-command publishing with npx agentplan upload ./plan.html
Stable slugs with immutable version history for every published document
Private, public, and password-protected visibility controlled by the owner
Hostile HTML isolated in a sandboxed iframe and stored in a private R2 bucket
Scoped API tokens stored as hashes for agent and CLI workflows
Admin controls, account quotas, moderation, and audit events
Insights
The public AgentPlan entry point and its one-command publishing workflow
Technical frame
The technical frame sits at the end on purpose: it shows the engineering substance behind the solution.
Next project
Measured, not marketed.
View project