OpenCode ChatGPT/OpenAI Plus/Pro Support

[matt-oakes]

It seems like currently OpenAI is only supported through OpenCode if it is using an API key. Would it be possible to also support it when using the Plus or Pro subscriptions?

[matt-oakes]

For reference, this works by adding this to the OpenCode config file:

  "provider": {
    "openai": {
      "options": {
        "baseURL": "http://localhost:8787/backend-api/codex"
      }
    }
  }

[mattpocock]

Triage: this is asking for a genuinely different auth path, not covered by the existing catalogue — OpenCode's openai provider entry here is API-key based, and ChatGPT Plus/Pro doesn't issue one. The workaround you posted routes through a local Codex-CLI backend proxy (http://localhost:8787/backend-api/codex), which is a different upstream shape than the request-logger proxy this tool builds around.

Flagging as needs-triage rather than picking it up directly, since it's a product-scope call: is subscription-based ChatGPT/Codex auth something we want a first-class catalogue entry for (similar to how codex's own chatgpt provider already works), or is the existing "Custom base URL" flow (pointed at your local Codex proxy) an acceptable answer for OpenCode specifically? @mattpocock — worth a decision before this becomes a ready-for-agent ticket.

[mattpocock]

Follow-up after digging into the code more: your workaround (baseURL: http://localhost:8787/backend-api/codex) isn't a coincidence — 8787 is this tool's own proxy port, and /backend-api/codex is the exact suffix Codex's existing ChatGPT subscription provider entry already uses (request-logger/agents.ts, codex → chatgpt, host chatgpt.com). So in principle this is "give OpenCode the same catalogue entry Codex already has," which sounded like a small, mechanical addition.

But there's a wrinkle: OpenCode's own openai provider entry carries an explicit warning already in the code —

"A ChatGPT login will not work here. OpenCode sends that traffic to a different host on purpose, so it goes around this tool. Use an OpenAI API key."

— i.e. a prior investigation (whoever wrote that note) concluded OpenCode's native ChatGPT-login flow ignores any configured baseURL for that session and always talks to the real chatgpt.com directly, unlike Codex which respects a -c openai_base_url=... override even for its ChatGPT-authenticated calls. If that's still accurate, a catalogue entry mirroring Codex's wouldn't actually work for OpenCode, and your workaround may have started successfully without the traffic actually round-tripping through the proxy.

@matt-oakes — could you confirm whether a real turn showed up in the request-logger's logs folder when you used that config, not just that OpenCode started without erroring? That would settle whether the existing warning is now stale (and this becomes a straightforward catalogue addition) or still correct (and this is blocked on OpenCode's own behavior, not something we can fix here).

[matt-oakes]

@mattpocock With the baseURL in the config file the request logger works correctly and I can see all of the log files. This points to the comment being state.