Which coding tools let you bring your own OpenAI-compatible endpoint

Tool by tool

What the endpoint has to get right

Speaking /v1/chat/completions is the easy half. The tools above also depend on details that "compatible" does not promise, and most reported failures come from these:

• Streaming tool calls. Arguments arrive as delta.tool_calls fragments and must reassemble into valid JSON. Every agent needs this.
• The Responses API for Codex, which only speaks it, including function_call_output replays and Codex's own namespace tools.
• Replaying history without reasoning_content. Most tools drop it; hosts that require it back fail the second turn (details).
• Unknown OpenAI-only fields. safety_identifier, store, metadata arrive from clients whether you want them or not (details).
• Tools with no parameters and forced tool_choice, which trips grammar compilers when {} becomes [] somewhere in the chain (details).
• A /v1/models list that includes every id you accept, because several tools validate the model name against it.

We test all of these against every chat model we serve and publish the results, failures included, on the model capabilities page.

Zed: the config the docs describe

{
  "language_models": {
    "openai_compatible": {
      "inferenceapis": {
        "api_url": "https://api.inferenceapis.com/v1",
        "available_models": [
          { "name": "deepseek-ai/DeepSeek-V4-Flash", "display_name": "DeepSeek V4 Flash", "max_tokens": 1000000,
            "capabilities": { "tools": true, "interleaved_reasoning": true, "parallel_tool_calls": true } },
          { "name": "zai-org/GLM-5.3", "display_name": "GLM 5.3", "max_tokens": 1000000,
            "capabilities": { "tools": true, "interleaved_reasoning": true } },
          { "name": "moonshotai/Kimi-K2.7-Code", "display_name": "Kimi K2.7 Code", "max_tokens": 262144,
            "capabilities": { "tools": true, "interleaved_reasoning": true } }
        ]
      }
    }
  }
}

Export INFERENCEAPIS_API_KEY before starting Zed; the variable name is derived from the provider key. interleaved_reasoning: true tells Zed to expect reasoning_content, which all three models return. We have not run Zed end to end ourselves yet; the fields are from Zed's documentation as of 2026-09-18.

Warp: what works and what does not

Warp added custom inference endpoints in May 2026 (Settings, search "inference endpoint"). The client sends your endpoint URL, key and prompt to Warp's backend, which makes the request, so two things follow: private and localhost URLs are refused, which is the open issue with the most votes, and Warp's own privacy terms apply to the prompt in transit. A public endpoint works. Ours takes the model ids from the model list; per-model settings such as context length are not configurable in Warp yet.

Claude Code

Claude Code talks to the Anthropic Messages API and nothing else. Pointing ANTHROPIC_BASE_URL at a plain OpenAI-compatible endpoint does not work; the host has to implement /v1/messages itself, including the streaming events, tool_use and tool_result blocks and thinking blocks. This endpoint does (https://api.inferenceapis.com/v1/messages), so the change is three environment variables. Claude Code 2.1.275 ran a fix-the-tests task end to end on eight models through it on 2026-09-18; the timings, costs and the gotchas are in the Claude Code guide.