All model inference clients disable HTTP redirects so prompts, credentials, and request bodies are not replayed to a redirect destination. This applies to every backend, not only Arcee. Extra headers are validated centrally for every backend and cannot override Host, Authorization, Proxy-Authorization, or x-api-key in any letter case; backend-selected credentials remain authoritative. Invalid header names and values are also rejected before dispatch.
Custom API-key endpoints are ordinary model configuration: any absolute public HTTPS base URL with a host is accepted without a separate host allowlist. Embedded URL userinfo is rejected. Plaintext HTTP is accepted by default only for loopback, private-network, and link-local destinations. A saved configuration or session may explicitly set allow_insecure_http = true to use a public HTTP endpoint; the single choice applies to its primary and optional light model and survives project inheritance, session updates, resume, and worker dispatch.
This opt-in is intentionally dangerous: the configured API key, prompts, source code, tool output, and model responses may be read or modified in transit. API-key requirements are unchanged, redirects remain disabled, credential-sensitive headers remain protected, and managed Arcee and ChatGPT Codex credentials keep their provider-specific issuer and origin binding. The flag bypasses only the public-HTTP transport rejection; it does not admit relative URLs, non-HTTP(S) schemes, missing hosts, embedded userinfo, or credentialless model configurations.