Support locally-hosted models via oMLX (or other OpenAI/Anthropic-compatible local servers) in Warp's embedded agents

[mriechers]

Problem

I run oMLX locally on Apple Silicon, which exposes OpenAI-compatible and Anthropic-compatible endpoints for locally-hosted models. I can already use these models like any other CLI tool in Warp's terminal (e.g., calling the endpoint directly, or via a CLI client) — that part works fine today.

The gap is that Warp's embedded agent features (Oz agent mode, AI command search, etc.) only support Warp's curated cloud model list, BYOK for direct provider APIs, or Enterprise BYOLLM (AWS Bedrock only). There's no way to configure a local endpoint like oMLX as the backend for those embedded features specifically.

Requested feature

Allow users to configure a custom local/self-hosted inference endpoint (e.g., oMLX's OpenAI- or Anthropic-compatible API) as a selectable model for Warp's embedded agent features, not just as a target for manual terminal commands.

Why this matters

• Privacy/offline use — keep embedded-agent inference local for sensitive codebases or offline work.
• Cost — no credit consumption when using local compute you already have running.
• Consistency — since oMLX already speaks OpenAI/Anthropic protocols, it should be a low-friction integration compared to a fully custom API.

Possible scope

• A "Custom/Local" model entry in Settings > Agents where a user supplies a base URL (and optional API key) for an OpenAI- or Anthropic-compatible endpoint.
• Could ship as experimental/opt-in, given tool-calling reliability may vary by local model.

[oz-for-oss]

@mriechers

This issue appears to overlap with existing issues:

• Support direct connections to localhost/private custom AI inference endpoints #12142 — Support direct connections to localhost/private custom AI inference endpoints
• Custom model: localhost / local URLs are rejected when adding a custom model #11589 — Custom model: localhost / local URLs are rejected when adding a custom model

If this report is meaningfully different, please comment with the additional context or distinguishing behavior so a maintainer can review it. Otherwise, a maintainer may close it as a duplicate after review.

Maintainer details

I concluded that feature request for true local/private custom inference endpoint support in Warp's embedded agents, duplicating existing localhost/private endpoint requests..

Duplicate reasoning

• Support direct connections to localhost/private custom AI inference endpoints #12142: Requests the same direct desktop connection to localhost/private custom inference endpoints rather than routing through Warp's backend.
• Custom model: localhost / local URLs are rejected when adding a custom model #11589: Describes the same inability to save or use localhost/127.0.0.1 local OpenAI-compatible model servers as Warp custom models.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz

[leesa-ui-w]

Hey @mriechers, thanks for the detailed writeup, the oMLX/OpenAI-Anthropic-compatible angle is a great concrete example. This is the same underlying ask as #12142 (supporting local/localhost custom inference endpoints as the backend for embedded agents), so we're going to close this one out and track it over there. Feel free to hop on that thread and add your oMLX use case and an upvote so it's counted.