🔑
No credential copying
The user signs into WordPress normally. AgentPress does not ask them to paste application passwords or API secrets into another service.
Built for the WebMCP Challenge · Open source
WordPress already knows what you can do. Now your agent does too.
AgentPress lets ChatGPT help inside the WordPress session you're already using. No copied API keys. No application passwords. No second permission system. Your human identity remains the authority.
WebMCP Site Tools
WordPress Abilities
Server-side revalidation
Same workspace. Shared goals.
The agent helps. The human decides what ships.
V0.1 concept
The agent helps. The human decides what ships.
IdentityYour existing WordPress login
InterfaceTyped WebMCP Site Tools
ControlAgentPress policy + approvals
AuthorityWordPress capabilities remain final
The problem
Agents shouldn't get a skeleton key to your CMS.
Most integrations introduce another credential, another permission model, or a remote service with broad access. AgentPress works from the permissions WordPress already understands.
🛡
Authority can narrow, never expand
WordPress sets the maximum authority. AgentPress can allow, stage or block an action, but it cannot give the agent permissions the human does not have.
✓
Consequential work stays visible
Drafts can be safe and reversible. Publishing, live navigation and other consequential changes can be staged for explicit human approval.
The collaboration model
Human in the loop, without making the human do all the work.
AgentPress is a control layer between WebMCP and WordPress abilities. The agent can understand, plan and act, while the application remains responsible for identity, permissions and final execution.
Sign in normally
Open the site inside ChatGPT's browser and authenticate with the same WordPress account you already use.
Discover capabilities
WebMCP exposes safe, typed tools. The agent sees what the current user can read, draft, request or change.
Do reversible work
Read the site, understand structure, create drafts and organise content within the user's effective permissions.
Stage the risky bit
Consequential changes become visible proposals. The human approves, WordPress revalidates, then AgentPress executes and audits.
Interactive concept
Same WordPress. Same agent. Different human, different capability envelope.
Switch the role below. This is the core AgentPress thesis in one interaction: the agent never receives more authority than the logged-in person.
Effective AgentPress capabilities
// The browser discovers a typed Site Tool
navigator.modelContext.registerTool({
name: "agentpress/create-draft",
description: "Create content as draft only",
execute: async (input) => {
// WordPress permissions revalidated server-side
}
});
navigator.modelContext.registerTool({
name: "agentpress/create-draft",
description: "Create content as draft only",
execute: async (input) => {
// WordPress permissions revalidated server-side
}
});
Proposed Change
Approval required
PRIMARY NAVIGATION · AP-C-123
→
Agent staged navigation change · 10:41
Awaiting explicit human approval
Execution boundary: approval is intent, not authorisation. On approval, the server checks the current session, object state, WordPress capabilities, AgentPress policy and approval validity before changing live navigation.
Open source · Challenge build
A WordPress control layer for the agent-native web.
AgentPress is being built as a focused WebMCP challenge project: real ChatGPT, real WordPress permissions, visible approvals and a complete audit trail.