Use case
Onlook already understands a project's styles and exposes internal agent capabilities such as style-guide reading. External coding agents working in the same repository need a compact, read-only version of that context before they edit UI code.
Proposal
Expose a vendor-neutral local project design-context surface that returns:
- semantic tokens and CSS variables
- mapped/reusable components
- typography, spacing, radius, and elevation conventions
- route/component evidence paths
- compact vs standard detail modes
- a stable schema with explicit byte bounds
The first implementation could be an internal tool/resource that later becomes part of the planned Onlook MCP surface. It should read current project state only and make no canvas or source mutations.
Why this is actionable
I have a working public proof of the consumer workflow in https://github.com/sarveshsea/chatbot and a deterministic Agent Skill at https://github.com/sarveshsea/remember-design-system. Memi currently builds this brief from source repositories, but an Onlook-owned context response would preserve Onlook's stronger visual and style knowledge.
If maintainers agree with the boundary, I can prepare a focused PR against the preferred internal agent-tool layer before touching MCP transport.
Use case
Onlook already understands a project's styles and exposes internal agent capabilities such as style-guide reading. External coding agents working in the same repository need a compact, read-only version of that context before they edit UI code.
Proposal
Expose a vendor-neutral local project design-context surface that returns:
The first implementation could be an internal tool/resource that later becomes part of the planned Onlook MCP surface. It should read current project state only and make no canvas or source mutations.
Why this is actionable
I have a working public proof of the consumer workflow in https://github.com/sarveshsea/chatbot and a deterministic Agent Skill at https://github.com/sarveshsea/remember-design-system. Memi currently builds this brief from source repositories, but an Onlook-owned context response would preserve Onlook's stronger visual and style knowledge.
If maintainers agree with the boundary, I can prepare a focused PR against the preferred internal agent-tool layer before touching MCP transport.