BYOK: Support custom model providers (e.g., DeepSeek) beyond Anthropic/OpenAI/Google

[attila-lin]

Summary

Warp's BYOK (Bring Your Own Key) feature currently only supports Anthropic, OpenAI, and Google API keys. Users who want to use other model providers like DeepSeek have no way to configure a custom API endpoint and key — and any attempted configuration gets silently reverted to "auto" on restart.

Problem

DeepSeek and other open-weight / third-party models are increasingly popular for their cost-performance ratio. Warp only allows BYOK for three providers, with no "custom provider" option to add an arbitrary API-compatible endpoint. This locks users into a small set of models and prevents them from using models they already pay for via their own API keys.

Specific issue: even if a user manages to configure DeepSeek through some workaround, the setting does not persist across app restarts and always reverts to the "auto" model.

Desired workflow

1. Open Settings > API keys (or a new "Custom Providers" section)
2. Add a custom provider with: name (e.g. "DeepSeek"), base URL, API key, and optional model list
3. The custom model(s) appear in the model picker dropdown with a key icon (BYOK mode)
4. Selection persists across app restarts and is available for both local agent conversations and Oz Cloud Agents

Artifacts

None attached

Warp version

v0.2026.06.03.09.49.stable_01

Operating system

Windows (PowerShell 5.1)

[warp-for-oss]

@attila-lin

Here's what I found while triaging this issue:

Warp's current docs describe Custom inference endpoints for OpenAI-compatible APIs, but they are local-agent only; Cloud Agents still use Warp credits today. The fallback-to-auto behavior needs the exact setup path to determine whether this is a feature access issue or a separate persistence bug.

I have a few follow-up questions before I can narrow this down:

1. When you search Settings for inference endpoint on v0.2026.06.03.09.49.stable_01, do you see a Custom inference / Add custom model section?
2. What exact workaround or configuration path did you use before the model selection reverted to auto after restart?

Reply in-thread with those details and the triage workflow will automatically re-evaluate the issue and update the diagnosis, labels, and next steps.

Maintainer details

I concluded that feature request for broader BYOK/custom model provider support, with an under-specified persistence symptom when attempting to use DeepSeek or another custom provider..

Triage summary

This is primarily an enhancement request for broader custom model-provider support beyond the fixed BYOK provider fields, with a secondary, under-specified persistence symptom where a manually configured DeepSeek/custom model reportedly falls back to auto after restart.

Current code and docs show two related but distinct paths:

• BYOK API keys are represented as fixed provider fields for OpenAI, Anthropic, Google, and OpenRouter-style provider keys.
• Custom OpenAI-compatible endpoints are implemented separately behind FeatureFlag::CustomInferenceEndpoints plus account/workspace eligibility checks, then surfaced as synthetic model choices from locally stored endpoint config.

That makes the likely issue either an access/rollout/eligibility gap for the custom inference endpoint UI on this reporter's stable Windows build, a product request to expand custom endpoint support to Cloud Agents, or a separate persistence bug in how a manually configured custom model/profile selection survives restart. The report does not include enough detail about the workaround used to distinguish those cases.

Duplicate check

I searched all open non-PR issues and recent closed issues for BYOK, custom inference, custom endpoint, DeepSeek, OpenAI-compatible, and provider-related requests. Related issues exist for local/private custom endpoint URLs, /models import, Anthropic-compatible endpoints, custom endpoint display issues, and DeepSeek thinking-mode runtime failures, but I did not find two likely duplicates of this broader provider-support + persistence + Cloud Agent request.

Suggested routing

Route to Agent/settings owners. The Cloud Agent portion likely needs product/backend consideration because custom endpoint credentials are locally stored today and are not available to cloud-hosted runs.

Question reasoning

1. When you search Settings for inference endpoint on v0.2026.06.03.09.49.stable_01, do you see a Custom inference / Add custom model section? — This determines whether the reporter lacks access to the intended custom endpoint UI on their account/build, versus seeing the UI but hitting a configuration or persistence bug.
2. What exact workaround or configuration path did you use before the model selection reverted to auto after restart? — The persistence symptom cannot be reproduced from the report without knowing whether the reporter edited a profile/model setting, used a hidden/custom endpoint UI, or changed local configuration manually.

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

Powered by Oz

[FuturMix]

Until Warp adds native provider extensibility, you can get custom model access today by routing through an OpenAI-compatible API gateway:

How it works with Warp's existing BYOK:

1. Set up a gateway that aggregates multiple providers (DeepSeek, Claude, Gemini, Llama, Mistral, etc.) behind a single OpenAI-compatible endpoint
2. In Warp's AI settings, configure the gateway as your custom OpenAI endpoint — one base URL + one API key
3. Warp sends requests to the gateway, which routes to the correct upstream provider based on the model ID

This sidesteps the provider lock-in without any Warp code changes. The gateway handles:

• Provider-specific auth — each upstream provider has its own API key, but you only manage one key in Warp
• Model catalog — /v1/models returns all available models across all providers
• Format normalization — different providers have different SSE streaming formats, response structures, and error codes. The gateway normalizes everything to standard OpenAI format before it reaches Warp

For the settings persistence issue: If you're pointing at a gateway, the base URL and key are the only two things that need to persist — no per-provider key management. So even with the current settings-reset bug, it's just two values to re-enter.

For custom model lists: Most gateways expose /v1/models which Warp could query to populate the model dropdown dynamically. This would make the model list self-updating as providers add new models.

Checklist for setting up OpenAI-compatible endpoint integrations (model ID formats, auth header passthrough, streaming SSE differences): https://futurmix.ai/docs/integration-guide?utm_source=github&utm_medium=issue_reply&utm_campaign=warp

Disclosure: I work on FuturMix, an OpenAI-compatible API gateway, and put together these docs while debugging similar multi-provider integration issues. Sharing in case it helps.

[esafak]

@attila-lin I think @oz-agent might be waiting for your response:

I have a few follow-up questions before I can narrow this down:

1. When you search Settings for inference endpoint on v0.2026.06.03.09.49.stable_01, do you see a Custom inference / Add custom model section?
2. What exact workaround or configuration path did you use before the model selection reverted to auto after restart?

Reply in-thread with those details and the triage workflow will automatically re-evaluate the issue and update the diagnosis, labels, and next steps.

[warp-for-oss]

@esafak

Here's what I found while triaging this issue:

Warp has a Custom inference endpoint flow for OpenAI-compatible endpoints, but custom endpoints are currently local-agent only and do not apply to Cloud Agents. The remaining uncertainty is whether your build/account exposes that flow and which configuration path caused the selection to reset to auto.

I have a few follow-up questions before I can narrow this down:

1. In Settings on v0.2026.06.03.09.49.stable_01, if you search for inference endpoint, do you see the Custom inference / + Add custom model section?
2. Which exact workaround or configuration path did you use before the model selection reverted to auto after restart?

Reply in-thread with those details and the triage workflow will automatically re-evaluate the issue and update the diagnosis, labels, and next steps.

Maintainer details

I concluded that warp already has local custom inference endpoint support, but the reported absence/reset behavior remains underspecified and cloud-agent support is a documented product gap..

Triage summary

This looks like an enhancement request with an underspecified persistence symptom mixed in. Current code and docs show that Warp supports Custom inference endpoints for OpenAI-compatible local agent inference, including user-entered endpoint URL, API key, and model IDs. However, that flow is gated by feature/entitlement checks and the docs explicitly say custom inference endpoints do not apply to Cloud Agents because the endpoint configuration is stored locally.

The reported auto reset still needs reporter detail before treating it as a confirmed bug. In current code, custom endpoint models are persisted in secure storage, surfaced as synthetic LLMInfo choices, and preserved by preference reconciliation when custom inference is enabled and the custom model is still valid. Custom selections can be cleared when custom inference is unavailable/disabled or the selected custom model can no longer be resolved.

Relevant implementation areas:

• app/src/settings_view/ai_page.rs renders the API Keys / Custom inference settings widget and gates the + Add custom model flow.
• app/src/settings_view/custom_inference_modal.rs validates endpoint URL/API key/model inputs.
• crates/ai/src/api_keys.rs persists provider keys and custom endpoints, and builds the custom model provider payload for requests.
• app/src/ai/llms.rs builds synthetic custom model choices and reconciles persisted model selections.
• app/src/ai/agent/api.rs includes custom model providers in local agent requests when the feature and workspace capability are enabled.
• docs/src/content/docs/agent-platform/inference/custom-inference-endpoint.mdx documents OpenAI-compatible custom endpoints and the current Cloud Agent limitation.

Duplicate check: I enumerated all open issues and checked recent closed candidates. The closest open issues (#11514 and #11586) are about auto-discovering models from /v1/models, and #12142 is about private/localhost endpoint reachability. They overlap with adjacent custom inference improvements but do not fully duplicate this report's combination of provider support, UI availability, cloud-agent scope, and reset-to-auto behavior.

Question reasoning

1. In Settings on v0.2026.06.03.09.49.stable_01, if you search for inference endpoint, do you see the Custom inference / + Add custom model section? — This determines whether the issue is about a missing gated settings surface versus behavior within the supported custom inference flow.
2. Which exact workaround or configuration path did you use before the model selection reverted to auto after restart? — The current code preserves valid custom endpoint model selections, so reproducing the reset depends on knowing whether the reporter used the custom inference modal, an OpenAI-compatible gateway/base URL path, edited preferences, or another path.

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

Powered by Oz