Agent Connector Canvas

Problem

When images, video, audio, and text prompts are placed in an AI creative canvas, the hard part is often not “can it generate,” but whether the nodes form an executable relationship: missing assets, matching inputs and outputs, Provider availability, and whether parameters reach a runnable state. Agent Connector Canvas addresses this engineering gap by observing the current canvas and asset state, then forming a minimal understandable workflow skeleton instead of dumping a chain of generation commands.

How It Works

• State observation: checks the current canvas, assets, prompts, and existing nodes to identify the user goal and gaps.
• Skeleton generation: produces a minimal understandable workflow, preferably as a Workflow Ready card or a workflow JSON draft.
• Capability validation: inspects each node’s inputs, outputs, and Provider capabilities, avoiding fake assets, placeholder images, blank images, or private local paths as deliverables.
• Local Provider setup: uses the API key entry in the right-side agent card to detect and save local configuration, such as DeepSeek API and Codex CLI; API keys are written only to .data/local-provider-config.json and are not echoed or committed to the repo.
• Safety boundary: does not read secrets, cookies, or browser login state; does not call external models directly, bypassing Provider; real paid generation requires an explicit switch and manual confirmation.

Boundaries

It is useful for organizing AI creative pipelines, generating validated workflow drafts, and configuring local Providers. It is not intended to automatically trigger batch real generation. If key assets are missing, it asks at most one required question. When using Codex CLI, the WindowsApps desktop app is not the CLI; verify that codex --version works, or use codex.cmd / codex on PATH.