# Build the Col MCP server and agent plugin · Issue \#40 · screen/col

[View on GitCafe](https://git.cafe/screen/col/issues/40)

Repository: [screen/col](https://git.cafe/screen/col)

Visibility: public

State: completed

Priority: 0

## Description

[Discussion and activity](https://git.cafe/screen/col/issues/40)

---

> Drafted with AI from the Col MCP research notes.

## Current state

Col already has library discovery, verified component links, and setup guidance, but coding agents can only reach it through the website. #4 covers listing third-party MCP servers in the catalog; this issue is about Col's own MCP server and agent plugin.

## Goal

Help coding agents choose the right UI component across libraries, with source-linked setup, dependencies, and variants. The host agent works in the developer's project; Col supplies the evidence.

## Workflow

1. **Find a fit.** For example: "Find an accessible React command menu for our Tailwind setup." Return a short list with framework, styling, dependencies, source links, and known gaps.
2. **Inspect the exact component.** Return the documented installation, API, example, variant, and source revision. Flag anything unverified.
3. **Implement in context.** The coding host inspects the project and follows the documented setup.

## Build order

These are acceptance gates, not a calendar.

1. **Catalogue MCP.** Expose the existing library search and verified component links through one remote HTTPS endpoint. Pass when IDs, filters, partial coverage, and the target clients work correctly.
2. **Documentation retrieval.** Cover about eight libraries in depth with setup, API evidence, and snapshots. Pass when reuse permissions are recorded, commands are verified, and an evaluation baseline exists.
3. **Compatibility retrieval.** Resolve versions and variants; return dependencies and supported constraints. Pass when hard filters are correct and unknowns are explicit in labelled tests.
4. **Plugin.** A thin package (`plugin.json`, `mcp.json`, and a `ui-components` skill) with one connection and the workflow above.
5. **Visual gallery**, after the above.

## First version

- Reuse the existing Next.js + TypeScript stack and catalogue IDs.
- One remote Streamable HTTP endpoint using the official TypeScript MCP SDK.
- Validated files for the small corpus; one Postgres service only once it grows.
- Exact IDs, hard filters, names, aliases, and lexical search. Add embeddings or reranking only if measured query failures justify them.
- Never request or store credentials.

## Open decisions

- Can the documentation, code, and preview media be reused? Record permission before ingesting.
- Which eight libraries get deep coverage?
- Which versions and variants does Col support? Keep unknowns explicit rather than substituting.
- Which client combinations actually work? Record host and protocol versions from deployed tests.
- Does Col improve real component tasks? Run a small pilot before expanding ingestion or the gallery.

## Acceptance criteria

- An agent can search Col's catalogue and get verified component links over MCP in the target clients.
- Responses link to their sources and mark unverified or partial coverage.
- No credentials are requested, stored, or committed.
- The plugin installs with one connection and documents the workflow.

