Add first-class AgentRouter provider once AgentRouter enables ForgeCode client support

[chindris-mihai-alexandru]

Summary

I prepared a local prototype for first-class AgentRouter support in ForgeCode.

The idea is a single-token agent_router provider with official endpoint behavior:

• default OpenAI-compatible route:
  • https://agentrouter.org/v1/chat/completions
• model listing:
  • https://agentrouter.org/v1/models
• Claude model IDs automatically rerouted to Anthropic-compatible:
  • https://agentrouter.org/v1/messages

This gives a simple UX for users who want one provider entry for:

• Claude models
• GLM/Zhipu models
• DeepSeek models
• other OpenAI-compatible models exposed by AgentRouter

Why this is useful

AgentRouter publicly documents support for:

• Claude Code
• Codex
• Gemini CLI
• RooCode
• Kilocode
• Qwen Code
• Droid CLI

So there is a reasonable case for first-class ForgeCode support as well.

Current blocker

AgentRouter currently rejects ForgeCode/generic client requests with:

unauthorized client detected

I reproduced that against:

• /v1/models
• /v1/chat/completions
• /v1/messages

However, with the same token, the official Claude Code client gets past client auth and reaches quota/model-entitlement errors instead.

That strongly suggests the blocker is on AgentRouter’s side:

• ForgeCode is not currently an allowed client
• official Claude Code is allowed

I opened an AgentRouter issue for that here:

• Support ForgeCode as an official client (currently rejected with "unauthorized client detected") agentrouter-org/docs#21

Local groundwork

I already implemented and committed local groundwork:

• commit: a5d418dff
• message: feat(provider): add AgentRouter routing groundwork

The local patch includes:

• ProviderId::AGENT_ROUTER
• provider config in provider.json
• routing logic for Claude vs non-Claude AgentRouter models
• tests for provider config and routing behavior

Suggested upstream path

I think the best next step is:

1. ask AgentRouter to officially support / whitelist ForgeCode
2. once confirmed, open a PR with the existing groundwork cleaned up for upstream review

Notes

This is different from just using generic custom providers because the AgentRouter UX is better as first-class support:

• one login flow
• one token
• one provider entry
• built-in Claude vs OpenAI route handling

If maintainers are open to it, I can convert the local commit into a proper upstream PR once AgentRouter confirms support.

Co-Authored-By: ForgeCode noreply@forgecode.dev

[chindris-mihai-alexandru]

Follow-up after broader AgentRouter client validation:

• ForgeCode / generic requests are still blocked by AgentRouter with unauthorized client detected on:
  • /v1/models
  • /v1/chat/completions
  • /v1/messages
• Claude Code is now verified to work end-to-end with the same token on claude-opus-4-6.
• Codex is currently not viable because modern Codex requires /v1/responses, which AgentRouter does not expose; I opened agentrouter-org/docs#22 for the stale/broken docs.
• Qwen Code reaches AgentRouter but crashes after a Claude response with Cannot read properties of null (reading 'choices'); I opened agentrouter-org/docs#23.
• Droid CLI reaches first token, then crashes with null is not an object (evaluating 'H.usage'); I opened agentrouter-org/docs#24 and Factory-AI/factory#987.

Net: this still looks like ForgeCode should wait for explicit AgentRouter-side support / whitelisting before upstreaming a first-class provider. The local groundwork remains useful once that blocker is resolved.

Co-Authored-By: ForgeCode noreply@forgecode.dev