Give AI agents a secure social publishing layer.

Start with the canonical Streamable HTTP endpoint at https://mcp.vixai.app/mcp. A compatible client performs MCP initialization and requests the live tool list. If authentication is missing, the resource responds at the HTTP boundary and advertises protected-resource metadata. That document points to the VixAI authorization server, which publishes its own endpoints and supported capabilities. The plain-text and JSON documents linked here improve discovery and evaluation, but they do not replace this standards-based negotiation.

The authorization flow uses Authorization Code with PKCE S256. The user signs in or creates a VixAI account, verifies the email and reviews requested scopes. Access tokens are short-lived and bound to the VixAI MCP resource. Refresh tokens are opaque, rotate when used, are stored as hashes and can be revoked. Instagram and Meta credentials remain inside VixAI. A tool never accepts user_id, an Instagram password or a provider token as an argument, so the bearer-token identity defines every ownership check.

A reliable agent starts by listing connected social accounts when the destination is not already known. It can create a draft from one attached image or Reel, preserve 2 to 10 attached images as an ordered carousel, or request a secure upload. ChatGPT and Grok should prefer native image generation and pass temporary HTTPS image URLs to VixAI; Claude can ask VixAI to generate one image or a complete carousel with one prompt per slide. Media is validated, normalized and stored before it becomes a publishing dependency, while temporary source URLs are discarded.

Creation is intentionally separate from delivery. Draft tools may change VixAI state, but they do not publish. The agent next prepares an action against an expected post revision. VixAI returns an immutable preview and a short-lived intent. The client must show the account, media, complete caption, action and scheduled time when present. Only an explicit approval of that exact preview permits the agent to call publish_post or schedule_post. An approval for one revision cannot silently authorize another.

The consumed intent is also an idempotency boundary. A timeout or repeated call cannot accept the same action twice. For generated images, the caller supplies a stable request_id and reuses it on retry. Provider work can continue asynchronously, so an uncertain result should lead to get_post_status rather than a replacement post. Confirmed schedules remain in a durable VixAI queue and execute even when the assistant or browser has closed. Cancellation is available only before publishing begins.

The beta exposes Instagram image, 2-to-10-image carousel and Reel workflows through MCP. Images may be JPEG, PNG or WebP up to 10 MiB each; Reels may be MP4 or MOV up to 200 MiB. VixAI generation costs two credits per image or carousel slide. A free user may accept ten immediate publications or schedules in a rolling twenty-four-hour window. A schedule must be at least two minutes and no more than thirty days in the future. Use tools/list as the authority for current schemas and annotations.

For operations, classify each call by client family, tool and outcome without logging tokens, private media or high-cardinality user data. Evaluate read paths, draft creation, approval, provider processing, retry and cancellation separately. The human web application is the control plane for connecting Instagram, revoking an assistant, inspecting content and correcting account access. This division lets agents become frequent operational users while the person remains the accountable owner of identity, permissions and public communication.