Build1 publisher2 min readPublished
VS Code 1.122 adds native Custom Endpoint provider for bring-your-own-key model routing via chatLanguageModels.json
The stable build includes a Custom Endpoint provider that speaks Chat Completions, Responses and Messages, so a compatible API and your own key are enough to get streamed chat and tool calls inside the editor.
The Engineer · Build desk
What happened
- A DEVUP AI guide reports that VS Code's native Custom Endpoint provider reached the stable build in version 1.122, supporting the Chat Completions, Responses and Messages API formats.
- The entry point is the Command Palette command "Chat: Manage Language Models", or the model picker in the Chat view, where Add Models offers Custom Endpoint as a provider type.
- For developers in Algeria, the guide's pitch is streamed chat and tool-capable workflows inside VS Code paid for locally in Algerian dinars, against a single public endpoint.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Which company holds the billing relationship for editor AI becomes a config decision inside VS Code, so a team can keep the editor and pay a model provider directly in its own currency.
- exposure A long-lived provider key now sits behind a reference in a JSON file that developers commit. The secure reference is the only thing keeping the raw secret out of the repository.
- constraint By the guide's own account bring-your-own-key covers chat and tools, so anyone costing out a subscription cancellation has to verify the remaining editor AI features feature by feature first.
- precedent With key plumbing and format selection in the editor itself, a routing extension has to justify its install on something other than holding the API key.
The wizard does not discover your models. Choosing Custom Endpoint in the Language Models editor asks for a group name, a display name and your key [8]. VS Code then opens `chatLanguageModels.json`, with the provider-level configuration and a secure reference to that key already written. The guide is explicit that you should not replace the reference with the raw value [9]. The models array is hand-edited: an exact model id, a display name, the endpoint url, and `toolCalling` [10]. The template's placeholders are intentionally not valid final values, and each has to be replaced with information from the model's catalog page before you save [11].
The prerequisite list shows where the work actually sits. Before you begin, the guide wants the exact model identifier from the catalog and the model's context and output limits in hand [12]. If you intend to run agent workflows, it also wants a model that explicitly supports tool calling [13]. You gather those yourself.
To run the same setup against some other provider, the endpoint must speak one of the three supported formats [2]. The catalog must publish the exact model id and the limits [12]. And the model has to do tool calling if you want agent use [13]. This walkthrough exercises Chat Completions [21]; Responses and Messages are listed as supported and not demonstrated [20].
The key hygiene advice is a dedicated key per environment, rotatable and revocable on its own. Then the specific prohibitions: no key in a screenshot, a Git repository, source code, a tutorial, an issue report or a screen recording [16].
The guide's claim for bring-your-own-key is bounded on its own terms. It says the experience "can power chat and tools without requiring a bundled AI subscription", and that this "does not automatically replace every AI capability inside the editor" [5]. A boundaries section is promised, and the available text breaks off before reaching it [6].
All of this comes from a DEVUP AI walkthrough that discloses it is independent of VS Code's publisher and not endorsed by it [18]. The same guide tells readers to verify current behaviour against the official language-model documentation and the 1.122 release notes [17]. Take the version number seriously for a second reason: the guide says a missing Custom Endpoint option means checking the installed build first [19].
What to watch
- Whether the official VS Code 1.122 release notes and language-model documentation enumerate the Custom Endpoint limits the guide defers.
- Whether anyone publishes a working Custom Endpoint config against a Responses or Messages endpoint, which the guide does not demonstrate.
- Whether inline completions and other bundled editor AI features function under a bring-your-own-key provider.