Codex Desktop with third-party APIs while keeping plugin/account compatibility

[dengmengmian]

AgentGate is a local AI gateway for coding agents. The newest documentation pass highlights a capability that is easy to miss but important for Codex Desktop users:

Codex Desktop can use third-party APIs such as DeepSeek and Xiaomi MiMo while keeping the official OpenAI-authenticated provider path that plugin and account features expect.

Why this matters

Many local proxies make Codex behave like a generic OpenAI-compatible client. That can be enough for plain model calls, but Codex Desktop also has account, provider, and plugin assumptions tied to its official OpenAI provider path.

AgentGate keeps that shape:

Codex Desktop
  -> OpenAI provider entry in Codex config
  -> local AgentGate base URL
  -> DeepSeek / Xiaomi MiMo / OpenAI / Kimi / GLM / DashScope / other provider

The goal is not just "make Codex call another model". The goal is to let users switch model backends while preserving as much of the native Codex Desktop experience as possible.

Useful docs

• Codex Desktop + third-party APIs + plugins: https://github.com/dengmengmian/agentgate-ai/blob/main/docs/use-codex-desktop-with-third-party-api-and-plugins.md
• Codex + DeepSeek: https://github.com/dengmengmian/agentgate-ai/blob/main/docs/use-codex-with-deepseek.md
• Codex + Xiaomi MiMo: https://github.com/dengmengmian/agentgate-ai/blob/main/docs/use-codex-with-mimo.md

Feedback wanted

If you use Codex Desktop with DeepSeek, MiMo, Kimi, GLM, DashScope, or another third-party provider, I would love to know:

• Which provider/model are you trying to use?
• Which Codex Desktop plugin/account features matter most to you?
• Which setup step is still confusing?
• What should be documented with screenshots or a short demo video first?

[jalu6868-boop]

这个项目做的非常好。从 0.9.0 开始使用的。

[dengmengmian]

建议更新下。1.4.0 修复了相当多的 BUG。现在这个版本才算稳定版。

[richardchen874-sys]

This is a very useful direction for Codex Desktop users.

Keeping the official OpenAI-authenticated provider path while routing the actual model backend to DeepSeek / MiMo / Kimi / GLM is much more valuable than just making Codex call another OpenAI-compatible endpoint.

For real coding workflows, I think the key things to compare are:

1. Whether plugins and account features still behave normally
2. Streaming stability during long coding sessions
3. Tool-call / response format compatibility
4. Latency when routing through a local gateway
5. Cost visibility per session or per task
6. Easy fallback between providers when one endpoint fails

I’m especially interested in the DeepSeek / BytePlus / China-model API side for indie devs and small AI teams. This kind of gateway approach seems very relevant for lower-cost coding workflows.

Happy to test or exchange notes if useful — are you mainly looking for feedback on DeepSeek setup, provider compatibility, or Codex Desktop plugin/account behavior?

[richardchen874-sys]

Thanks for clarifying — that makes sense.

I can probably help with this testing. On my side, I already have access to DeepSeek / Kimi / GLM API backends, so I can test the provider compatibility part with real endpoints instead of only mock configs.

The areas I’d be most useful for testing are:

1. DeepSeek / Kimi / GLM routing through the local gateway
2. Streaming behavior during longer Codex Desktop sessions
3. Tool-call / response format compatibility
4. Provider fallback behavior
5. Cost visibility across different model backends

I’m especially interested in validating whether these China-model APIs can provide a lower-cost but still reliable Codex Desktop workflow for indie devs and small AI teams.

If helpful, I can set up a few test cases and share notes/logs back here. I can also provide some initial free testing quota for DeepSeek / Kimi / GLM endpoints so the integration can be tested more easily.

[dengmengmian]

Thanks, that would be very helpful.

DeepSeek / Kimi / GLM compatibility is exactly one of the main areas I want to validate, especially with real endpoints and real Codex Desktop workflows.

For safety, please don’t share any API keys or private credentials here. I think the best way is for you to test with your own provider accounts and share the results back in this discussion.

The most useful feedback would be:

1. OS and Codex Desktop version
2. Which provider/model was tested, for example DeepSeek / Kimi / GLM
3. Whether routing through the local gateway works correctly
4. Whether streaming remains stable during longer coding sessions
5. Whether tool calls and response formats work as expected
6. Whether provider fallback behaves correctly
7. Any latency or cost differences you notice
8. Logs, screenshots, or reproducible cases, with sensitive tokens removed

I’d especially like to know whether these China-model APIs can provide a lower-cost but still reliable Codex Desktop workflow for indie developers and small AI teams.

Thanks again — real-world testing feedback like this would be very valuable for improving the project.

[richardchen874-sys]

That’s great—the eight points you mentioned are crucial to us. Thank you for the clear list!

Once our independent site completes its internal beta testing, I will provide you with some free credits for testing purposes. User feedback—and whether we successfully solve real user problems—will determine the scope of what we can achieve.

If you could provide your email address, I will send the API key to you in three days.

[PwccaCode]

I also have access to glm, kimi, deepseek, qwen and mimo

will run some tests using free openai account xD

On windows (always some installation issue)

[dengmengmian]

I also have access to glm, kimi, deepseek, qwen and mimo

will run some tests using free openai account xD

On windows (always some installation issue)

Thanks! Windows testing would be especially helpful.

Could you please share a bit more detail about the installation issue you mentioned?

It would help a lot if you can include:

1. Windows version, for example Windows 10 / 11
2. AgentGate version / release package you tested
3. Codex Desktop version
4. How you installed it, for example .exe, .msi, or portable build
5. The exact error message or screenshot
6. Whether the local gateway starts successfully
7. Whether the Codex Desktop config is written correctly
8. Which provider/model you tested, for example GLM / Kimi / DeepSeek / Qwen / MiMo

If you can, please also attach logs with any API keys or private tokens removed.

Windows installation and first-time setup are areas I really want to improve, so this feedback would be very valuable.