Using DeepSeek models with Codex CLI.

Yes, what you're seeing is expected based on the capabilities currently exposed by the Azure hosted DeepSeek models.

A key distinction is that Codex CLI is designed around tool-enabled/agentic workflows, not just text generation. Beyond producing responses, Codex relies on capabilities such as tool calling, command execution, file operations, and MCP-style integrations.

According to the Azure AI Foundry model capability matrix, DeepSeek-V4-Pro currently does not advertise tool calling support in Azure-hosted deployments. As a result, even if the model itself can generate code, it cannot provide the full set of capabilities that Codex CLI expects for agentic operations. This differs from GPT-5.x deployments that expose the required tool-enabled capabilities through Azure.

One subtle but important point is that "DeepSeek supports Codex" and "DeepSeek on Azure supports Codex CLI workflows" are not necessarily equivalent statements. Azure exposes a specific capability set for each hosted model, and Codex CLI compatibility ultimately depends on the capabilities available through the Azure endpoint rather than the vendor's standalone offering.

There is also an active discussion in the Codex ecosystem showing that DeepSeek models can sometimes be configured as custom providers, but tool/MCP-related functionality may not be available or may behave differently depending on the provider implementation. That aligns with the behaviour you're observing where standard GPT deployments work while DeepSeek-V4-Pro does not.

Bottom line: Based on the currently published Azure model capabilities, DeepSeek-V4-Pro is not a suitable choice for full Codex CLI agent workflows because the required tool-calling functionality is not available. If your workflow depends on file editing, terminal execution, MCP servers, or other Codex agent capabilities, use an Azure model that explicitly supports those features.

Help make this community better for everyone: if this answer resolved your issue, please accept it or leave an upvote. If not, share more details in a comment so we can continue the discussion and find the right solution.

Azure deployments don't work out-of-the-box with Codex CLI because of header and URI mismatches (api-key vs Bearer authorization).

You can bypass this by setting up LiteLLM or Azure API Management as a proxy:

Spin up LiteLLM pointing to your Azure deployment URL.

Point your CLI's OPENAI_API_BASE to http://localhost:4000.

Verify your endpoint uses the Azure OpenAI-compatible route (.../models/chat/completions?api-version=...).