Only the latest tagged release of Talaria is actively supported by this document.

Talaria is a local-only bridge for Claude Code. It does not proxy or transform authentication secrets into third-party APIs. It expects an existing local codex authentication session and forwards requests to:

https://chatgpt.com/backend-api/codex/responses

• Does not accept OPENAI_API_KEY.
• Does not support Anthropic direct API passthrough.
• Does not support generic OpenAI-compatible providers.
• Does not support local model backends.

• Bound only to loopback hosts.
• Browser-origin requests are rejected on /v1/messages.
• Content-Type: application/json is required for message requests.
• Request bodies are capped at 8 MiB.
• Upstream response conversion is bounded.
• Gateway cache files are written with private permissions and symlink targets are rejected.
• Model IDs are sourced only from logged-in Codex visibility data.

Talaria v1 is intended for trusted local workstations. The gateway is loopback only and blocks browser-origin requests, but it does not yet require a separate gateway authentication token from non-browser local clients. Do not run Talaria on shared machines where other local users or untrusted local processes should not be able to connect to your loopback services.

Please open a private report with repository maintainers. Avoid posting sensitive credentials in public issue comments.

• Keep talaria installed from trusted sources only.
• Do not write or script around ~/.claude/cache/gateway-models.json.
• If TLS verification fails for chatgpt.com, resolve Python trust-store issues before running live model commands.