Cursor
Cursor is a closed-source IDE that supports a custom OpenAI base URL for its chat/composer surface. Point it at rapid-mlx and every model the server hosts becomes selectable inside the editor.
/v1/chat/completions (IDE chat/composer path) ·
Setup: Cursor Settings → Models → OpenAI Base URL ·
Matrix cell: no dedicated cell (Cursor's agent path is
locked to Cursor's backend — see gotcha below).
Tier-2 because only the IDE-chat surface is integrable and it needs the GUI app — the CLI/agent path can't point at a custom endpoint, so it can't get the per-release headless guarantee a Tier-1 agent gets.
Install
Download from cursor.com (closed-source vendor binary).
Config
Settings → Models → Add Model, then:
OpenAI API Base: http://localhost:8000/v1
API Key: not-needed
Model name: default # or any served alias (qwen3.5-9b-4bit, gemma-4-12b-4bit, …)
The default chat/composer flow honors this custom URL for OpenAI-shape endpoints. Cursor's agent/composer mode uses tool calls automatically — rapid-mlx handles them natively with tool-trained families (Qwen 3.5/3.6, Gemma 4, gpt-oss).
Run
$ rapid-mlx serve qwen3.6-35b-4bit
Then open Cursor and pick default (or the served
alias) from the model dropdown. All chat / composer requests now
route to http://localhost:8000/v1.
Gotchas
-
Cursor CLI's agent path routes exclusively through Cursor's
backend. Per Cursor's community forum + docs, the CLI/agent
path is not integrable at a custom endpoint (the IDE
chat/composer is). The 0.10.2 PR-2 pre-flight verdict recorded
this as DEFERRED — no cursor cell was added to
test_agents_matrix.py; qwen-code was promoted to the fallback slot instead. Verdict source:tests/integrations/README.md. -
Model dropdown may hide unknown aliases. Cursor validates
model names against its internal catalog; if a rapid-mlx alias
doesn't show, pick
defaultwhich always resolves to whatever the server booted.