Give your coding agent tool calls for searching the catalogue and fetching block source.
The manifest tells an agent what exists. The MCP server lets it act on that — searching the catalogue and pulling exact source without a web fetch.
It is a remote server, so there is nothing to install and no key to paste. Your client signs in to your Toro UI account with OAuth the first time it connects.
Add the URL to your MCP client configuration:
On first use your client opens a browser window to authorise. Access can be revoked per client from your account at any time.
Because config files get committed. OAuth keeps the credential out of your repository, scopes it to a single client, and lets you revoke it without rotating the key your CLI depends on.
| Tool | Licence needed | Purpose |
|---|---|---|
search_blocks | No | Find blocks by description, section or category |
list_sections | No | Enumerate sections and their block counts |
get_convention | No | Read a documented convention by topic |
get_block | Yes | Fetch a block's full source by id |
Search works without a licence on purpose. An agent should be able to tell you exactly which block solves your problem before you decide to buy one.
The agent never guessed a component name, and the result uses your accent colour because the block was already token-driven.
Without a licence that second call comes back as a prompt rather than a failure — the agent will describe the block and tell you where to buy it.
Source that an agent writes through MCP is not recorded in
toro.lock.json, so toro update will not know about it. If you
want a block to receive upstream improvements, install it with the CLI:
The tool surface may still change. search_blocks and get_block are
stable; the others may gain arguments.