What the agent can do

The agent's capabilities — what it can read, build, change, run, and inspect on your behalf. Everything that writes, runs, or deletes pauses at the side-effect confirmation gate for your approval before it executes.

Managing WordPress directly

On any install — no other plugin required — the agent reads and manages your content, taxonomies, media, users, comments, settings, menus and site overview, and works with popular plugins (WooCommerce, SEO, forms, LearnDash, MemberPress) when they are active. See Managing WordPress directly for the full list.

Workflows

The agent can list, read, design, modify, activate, deactivate, and delete workflows. Designing a new workflow or changing an existing one is a write action and always requires your approval before it lands.

Executions

The agent can browse past executions and inspect the full timeline of any one of them — every node visited, its inputs and outputs, and any errors. It can also trigger a new run, or replay a past execution from a specific step (useful for recovering from a downstream service that was temporarily down).

Connections

The agent can list stored credentials (values are masked) and add, update, or remove them. Adding or rotating a credential is a write action; removing one that is currently in use is refused unless you explicitly force it.

Records (with WP-PFManagement)

When WP-PFManagement is installed, the agent can list, query, read, and modify records of any entity it has permission for. Designing new entities or fields is a write action and requires your approval.

System inspection

• Health check — aggregate health of providers and the queue worker.
• Diagnostics dump — site fingerprint, plugin versions, recent error rates. Useful to attach to a support ticket.

Grounded answers

The agent can search this documentation site to ground its answers in real reference material rather than guessing capabilities. If you notice it claiming a feature that does not exist, that is the signal that the docs and the implementation drifted — please report it.

Why everything is typed

Every capability above is described to the LLM with an explicit input and output schema. When the LLM proposes an action, the proposal is checked against that schema before anything runs — invalid proposals are rejected with a structured error message the LLM uses to self-correct on the next turn. You should rarely see this; when you do, it shows up in the trace.