Can I use a local or self-hosted model?

Today: no. BYOK supports Anthropic, OpenAI, and Google only. There's no surface for pointing the playground / AI pair / evaluation judge at a local Ollama / llama.cpp / vLLM endpoint.

Workarounds

• Use Prompt Assay as a library of prompts, pull via the public REST API or SDK, and run them against your own infra. The resolved-content endpoint gives you the assembled prompt ready to send to any provider.
• Target-only model entries: Prompt Assay's model list includes entries for Llama 3.3 70B and Mistral Large. These are valid target models for metadata purposes (linter, token budget, export) but are NOT callable from the playground. They exist so you can manage prompts destined for self-hosted models.
• The target-only list is wider than self-hosted models. Four first-party models are also target-only, for reasons that have nothing to do with self-hosting: Claude Fable 5 (turns run many minutes to hours, which doesn't fit our interactive streaming panels), GPT-5.5 Pro and GPT-5.4 Pro (OpenAI serves them on the Responses API only, and our adapter speaks Chat Completions), and GPT-5.3 Codex (coding-specialised). All four can be set as a prompt's Target Model and get full linter, token-budget, and export support · they just aren't offered as a Workbench Model or in the playground.

Roadmap

We're watching demand for OpenAI-compatible endpoints as the canonical self-host protocol. If your org needs this today, email support. The shape of the feature is less interesting than how many teams have real production workloads gated on it.