ai coding agents don't have taste. you do (maybe).

A basic, stateless MCP that agents call when making decisions, backed by your curated development bookmarks.

So it's your fault now when things go wrong.

MCP

The MCP server is available over Streamable HTTP at:

http://localhost:8787/mcp

Point an MCP client or coding agent at that URL. The endpoint implements the final 2026-07-28 protocol and also accepts stateless legacy clients. Every request receives a fresh server instance: there is no session ID, sticky routing, or per-client memory.

Agents can use:

• consult_taste to search the library for a development decision;

• get_taste_item to inspect one exact bookmark;

• list_taste_categories to discover the available taxonomy;

• taste://library and taste://items/{itemId} as read-only resources;

• make_tasteful_decision as a reusable decision prompt.

The MCP surface is intentionally read-only. Curation and human review stay in the Web UI.

For access from another machine, expose the app and explicitly allow the hostname or IP used by MCP clients:

APP_HOST=0.0.0.0
MCP_ALLOWED_HOSTS=localhost,127.0.0.1,192.168.1.5

The .env file is optional. Without an API key, the enrichment flow runs in demo mode.

Related MCP server: bookmark-mcp

Configure AI enrichment

Add your DeepSeek API key to .env:

DEEPSEEK_API_KEY=your-key

• Docker binds the application to 127.0.0.1 by default.

• The application currently has no authentication. Do not bind it to a public interface unless it is protected by an authenticated HTTPS proxy.

• Public GitHub pages are read without a GitHub token. Private repositories are not supported.

Your bookmarks and reviews remain in your local SQLite database. Back up the Docker volume or the local data/ directory before moving installations.