Raycast Bridge

Use Raycast AI as a chat model inside VS Code, run AI quick actions on your selection, and fire any Raycast command from the editor.

Raycast exposes no API to other processes, so this extension pairs with a small companion Raycast extension that lives in the same repository. See Setup.

A quick action on the editor selection: the answer streams in underneath and says which model answered.

Features

Raycast AI in the Chat view. Raycast's models appear in the chat model picker under the Raycast AI vendor. Pick one and chat as with any other model. Raycast: Select Chat Models chooses which models are offered.

Ask AI. Raycast: Ask AI sends the selection, or the whole file, to the Chat view with an instruction you type.

Quick actions on the selection. Select text and a lightbulb entry and a code lens offer actions such as Translate to Chinese, Explain and Polish as Academic English. The answer streams in under the selection, or into a hover. Answers render as Markdown; give an action "render": false to show its answer verbatim instead, which suits rewrites of LaTeX or code where Markdown would eat characters or refuse to wrap. An inline answer takes keyboard focus when it appears, so Escape dismisses it; set raycastBridge.quickActionsFocus to false to keep focus in the editor. VS Code caps a comment thread at about 20 lines and does not refresh its scrollbar after that, so the first time an answer is longer the extension offers to turn off comments.maxHeight, which lets answers expand in full. Each action has its own prompt and can pin its own model with Raycast: Select Quick Action Model. Bind one to a key by writing a keybinding by hand:

{
  "key": "cmd+k t",
  "command": "raycastBridge.quickAction",
  "args": { "action": "Translate to Chinese" },
  "when": "editorTextFocus"
}

A key press with nothing selected runs the action on the line the cursor is on, so short edits need no selecting first. The lightbulb and the code lens still wait for a real selection, since either would otherwise offer itself on every line of every file.

Answers are cached. Running the same text through the same action and model replays the stored answer instead of spending Raycast AI quota again, which the per-minute and per-hour limits make worth doing. A replayed answer is labelled cached and carries a Regenerate button that asks Raycast afresh. The last 100 answers are kept; Raycast: Clear Quick Action Cache forgets them. Turn the whole thing off with raycastBridge.quickActionCache, or one action at a time by giving it "cache": false. Whitespace at the end of the selection is ignored when matching, since a drag easily overshoots into the newline; indentation at the start is not, since it is what an indented answer was written to match. Pinning a model keys the cache to that model; an action with no model pinned is keyed to the installed Raycast version instead, since that is what decides which model answers.

Run Raycast commands. Raycast: Run Command fires any deeplink listed in raycastBridge.commands, optionally injecting the selection.

Requirements

• macOS. Raycast is macOS only.
• Raycast 2.x with a Pro subscription for the AI features.
• The companion Raycast extension, installed once as described below.

Setup

1. Install this extension from the Marketplace.

2. Install the companion Raycast extension. It is not on the Raycast Store, so it is imported from source with the Raycast CLI:

   git clone https://github.com/youyinnn/vscode-raycast-bridge.git
   cd vscode-raycast-bridge/raycast
   npm install
   npx ray develop
   

   ray develop builds the extension, imports it into Raycast and then keeps watching. Once Raycast shows VS Code Bridge, press Ctrl+C. The extension stays installed.

3. Open the Chat view, pick a Raycast model, and send a message. The first launch shows a Raycast confirmation dialog. Choose Always so later launches are silent.

4. Optionally run Raycast: Probe Which Models Work. Raycast silently answers with its default model when it does not accept the id you asked for, and the probe finds out which ids the installed Raycast actually honours.

Commands

Settings

How it works

VS Code starts a loopback HTTP server, registers a job, and opens a raycast:// deeplink carrying only a job id, port and one-time token. The Raycast command fetches the prompt over HTTP, runs AI.ask, and streams the answer back. The server binds 127.0.0.1 on an ephemeral port, checks the token with a constant-time compare, rejects any request carrying an Origin header, and deletes jobs on completion.

Known limits

• Raycast AI quota is 10 requests per minute and 100 per hour. Every chat turn and quick action spends one.
• Raycast AI has no session, so the chat history is replayed into every request.
• Cancelling a request stops rendering, but Raycast keeps generating and still spends quota.
• The models also appear in Agent mode, where they do not work: AI.ask cannot call tools.
• Raycast cannot list installed extensions, so raycastBridge.commands is filled by hand.

License

MIT