Faceful MCP server setup
The shared studio and agent interface, with current availability made explicit.
The intended MCP workflow lets an existing Codex or Claude Code agent plan three avatar-video items, read their current revisions, revise selected content and prepare an exact schedule for owner authorization.
The shared engine, credential issuer and remote transport are pending integration. There is no verified OAuth authorization server or protected-resource metadata to advertise. A readable setup page is not a successful client handshake.
What the agent will operate
The proposed tool surface covers context, campaign creation and reading, version-checked revisions, bounded rendering, immutable schedule preparation, authorized scheduling, campaign discovery and cancellation. UI and MCP must use the same tenant-scoped domain service and current objects.
Private tools require explicit scope. Preparing a schedule does not authorize publication. The model cannot widen its own grant, supply its own actor identity or replace a human approval with a boolean.
Prepare a local client
The setup page has Codex and Claude Code command/config templates. Local CLI syntax was checked with Codex 0.169.0 and Claude Code 2.1.150; a fresh authenticated connection still needs to be tested after the backend is supplied. Other MCP hosts are not tested.
What works today
Call public status, read the public OpenAPI specification, or use the repository CLI to inspect documentation and validate a draft locally. No credentials are required for these reads.
Connection evidence and revocation
Connection success must come from a recorded authenticated context call with the connection name, scopes, expiry and last activity. Copying a command is not evidence. Revoking a token blocks new calls; stopping accepted jobs or cancelling a provider-accepted schedule requires a separate action with an actual per-job outcome.
First production acceptance test
After engine integration, test a fresh Codex client and a fresh Claude Code client: authenticate, discover tools, persist a campaign, inspect it in the studio, edit one item and read that exact revision back from the agent. Then verify a real render and each platform’s scheduling receipt. Record versions, elapsed setup time and failed steps before claiming a two-minute setup.