GLM Models for Raycast

A Raycast extension that exposes Z.ai / BigModel GLM models to Raycast's built-in AI features — AI Chat, Quick AI, and AI Commands — using Raycast's extension AI model provider API.

What it does

• Lists your account's GLM models (GLM-5.3, GLM-5.2, GLM-5.3-Flash, GLM-4.6, GLM-4.5 family, GLM-4.5V/4.6V vision models, …) in Raycast's model picker
• Streams completions straight from Z.ai's OpenAI-compatible API
• Supports reasoning (thinking + reasoning-effort picker), tool/function calling, image attachments on vision models, and system messages
• Works with both platforms: Z.ai international (api.z.ai) and BigModel China (open.bigmodel.cn) — switchable in preferences

Requirements

1. Raycast 1.26.0 or newer, signed in with your Raycast account

2. Raycast Pro — Raycast requires a Pro subscription to use models provided by extensions

3. Node.js 22.22.2+ and npm 7+ (check with node -v — matches @raycast/api's own requirement, also declared in engines)

4. An API key:

   • International: create one at z.ai → API Keys (works with the Z.ai (pay-as-you-go) platform setting)
   • China: create one at open.bigmodel.cn (works with the BigModel (pay-as-you-go) setting)

   Keys are not interchangeable between the two platforms. Pay-as-you-go API keys only — GLM Coding Plan and Team Plan keys are not supported.

Setup

cd glm-models
npm install
npm run dev

Running npm run dev registers the extension locally in Raycast (it appears at the top of root search).

First-run onboarding

The first time Raycast needs the extension's settings, it shows a setup form asking for every required preference together — API Key and Platform — with a help page (from help.md) beside the form. Pick the platform that matches where your key was created; the two platforms' keys are not interchangeable.

Picked "Custom base URL"? The setup form only collects the required preferences, so it never asks for the URL itself. After the form, set Custom Base URL in Raycast Settings → Extensions → GLM Models to your HTTPS OpenAI-compatible endpoint (e.g. https://open.bigmodel.cn/api/paas/v4), then run Refresh Models. Check Setup can validate a URL typed into it, but nothing typed there is ever saved.

Preference	What to enter
API Key	Your Z.ai or BigModel key
Platform	Z.ai (pay-as-you-go) or BigModel (pay-as-you-go) — must match where the key came from. Pay-as-you-go keys only: GLM Coding Plan / Team Plan keys are not supported
Custom Base URL	Only if you picked Custom (any HTTPS OpenAI-compatible endpoint) — not part of the setup form; set it in the extension settings
Extra Models	Optional comma-separated model IDs to force-include in the picker

After the setup form, validate everything with the Check Setup command: it calls the live API with your saved key and platform and tells you exactly what's wrong if they don't match (401 → key/platform mismatch), the endpoint is unreachable, or all good (N models discovered, listed below the result). It can also test a different key/platform combination before you commit it to preferences — a combination that differs from the saved one is reported as validated, but not applied, since extensions cannot change their own preferences — and its Open Platform Console action deep-links to the selected platform's API-keys page (z.ai or open.bigmodel.cn). Validating the saved combination also refreshes Raycast's model list.

Then opt in to extension models (one-time):

1. Open Raycast Settings → AI (or the model picker in AI Chat)
2. Enable models from extensions ("Allow AI Models" / toggle this extension on)
3. Open AI Chat or Quick AI, open the model picker, and pick a GLM model — glm-4.5-flash is free, a good first smoke test

While developing, keep npm run dev running for hot reload. Press ⌃C to stop; the extension stays in Raycast. Re-run npm run dev after code changes.

The Refresh Models command validates the saved key + platform and re-runs model discovery in one step, reporting the outcome (e.g. "Z.ai (pay-as-you-go): 18 models available") — the one-step way to pick up a platform change from the settings (Raycast also refreshes automatically in the background, but it doesn't notify extensions when preferences change).

The Show Models command lists everything the extension provides to Raycast's model picker in a master-detail view: the searchable model list on the left, and the selected model's full metadata on the right — display title, model ID, context window, vision/reasoning/tools capabilities, whether it comes from the Extra Models preference, and where the list came from (the platform's live /models endpoint, or the models.dev / curated fallback when that is unreachable). Run it after changing preferences to see exactly what Raycast will receive.

How model discovery works

• The extension first calls GET {baseURL}/models with your key to get the models your account can actually use (handles platform differences and new releases automatically). The result is cached for 60 seconds in memory and mirrored to disk, so Raycast's frequent background polling doesn't re-fetch the endpoint.
• Each ID is enriched with metadata (title, context window, vision/tools/reasoning capabilities) from the models.dev community catalog, cached for 24h — BigModel reuses the Z.ai catalog since both platforms serve the same model IDs.
• IDs models.dev doesn't know get conservative defaults, so brand-new GLM models still show up and work.
• If the /models call fails, the models.dev model list is used instead; if that is unreachable too (e.g. first run offline), a small curated fallback keeps the picker populated: GLM-5.3, GLM-5.2, GLM-5.3-Flash, GLM-4.7, GLM-4.6, GLM-4.5-Flash.

Notes:

• Reasoning effort maps to Z.ai's reasoning_effort on GLM-5.x models, and to thinking on/off on GLM-4.x models (none disables thinking).
• Reasoning tokens stream into Raycast's collapsible thinking section (reasoning-delta parts).
• Image attachments are only delivered for models declared with vision (GLM-*-V and Flash multimodal models).

Troubleshooting

• Run "Check Setup" first — it validates your saved key and platform against the live API and classifies the failure (key/platform mismatch, network, endpoint unavailable).
• Picked "Custom base URL" during setup and no models appear — Raycast's first-run setup form only collects the required preferences (API Key, Platform), so the URL is still unset. Set Custom Base URL in the extension settings, then run Refresh Models. Check Setup only validates — a URL typed there is never saved.
• Changed Platform but the model picker is stale — Raycast doesn't notify extensions when preferences change. Run Refresh Models to validate and refresh in one step. Note that Check Setup only validates: after testing a different platform or key, save it in the extension settings for it to take effect.
• AI_DownloadError: Cannot find module 'undici' when attaching an image — this happened in early versions of this extension: image attachments were converted to data: URLs, which the AI SDK tries to fetch, and that path requires the undici npm package which Raycast's runtime doesn't provide. Fixed by passing image bytes directly; if you see it again, run npm run dev so you're on the latest build.
• No GLM models in the picker — check the npm run dev console: it logs either glm-models: discovered N model ids via …/models (dynamic discovery worked) or glm-models: /models lookup failed … followed by which fallback list was used (models.dev or the curated catalog). Also confirm you toggled the extension on in Raycast Settings → AI.
• 401/403 errors — your API key doesn't match the selected Platform (z.ai vs BigModel keys are not interchangeable).

Publishing to the Raycast Store (optional, later)

This repo is currently a personal extension. If you want to publish it:

1. Have a GitHub account linked to Raycast (npm run publish authenticates with GitHub)
2. npm run build to validate for distribution
3. npm run publish — opens a pull request against the raycast/extensions repository
4. After Raycast team review and merge, the extension is published to the Store automatically

Development

src/models.ts           AI model provider entry point: getModels + streamCompletion
src/lib/catalog.ts      Curated model metadata, /models discovery, preferences
src/lib/format.ts       Shared model display formatting (context window, capabilities)
src/check-setup.tsx     Command: validate key + platform against the live API
src/refresh-models.ts   Command: manually refresh the model list
src/show-models.tsx     Command: browse the models provided to Raycast

• npm run dev — run in development mode with hot reload
• npm run build — production build / store validation
• npx tsc --noEmit — type check

The provider uses @ai-sdk/openai-compatible + Vercel AI SDK streamText, pointed at Z.ai's OpenAI-compatible endpoint. Provider options under the zai namespace (thinking, reasoning_effort) are forwarded into the request body by the compatible provider.