Coding agents

Any tool that speaks the OpenAI API works here: paste the base URL, an API key and a model ID and you are running. What most tools cannot discover on their own is each model’s real limits — this page gives you a config that fills them in, plus the three settings that prevent the most confusing failures we see in real logs.

The two-minute setup

Every tool asks for the same three things: base URL https://api.clfaigateway.dev/v1, an API key from your dashboard (starts with sk-gw-, sent as a standard Authorization: Bearer header), and a model ID from the table below. All nine models support tool calling and reasoning.

Valid reasoning_effort values differ per model — send one the model does not accept and the error message lists the right set. Live pricing for all nine is on the models page.

opencode: fill in what the form cannot

The Custom provider form in opencode only stores model IDs and display names, so every model shows Context limit 0 and no reasoning support — which breaks auto-compaction and hides thinking. Paste this into ~/.config/opencode/opencode.jsonc instead (or merge the provider block into your existing file), then restart opencode:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "clf-gateway": {
      "name": "CLF AI Gateway",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://api.clfaigateway.dev/v1"
      },
      "models": {
        "glm-5.3": {
          "name": "GLM 5.3",
          "attachment": false,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 1048576, "output": 32768 }
        },
        "glm-5.3-flash": {
          "name": "GLM 5.3 Flash",
          "attachment": true,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 1048576, "output": 32768 }
        },
        "glm-5.2": {
          "name": "GLM 5.2",
          "attachment": false,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 262144, "output": 32768 }
        },
        "glm-4.7-flash": {
          "name": "GLM 4.7 Flash",
          "attachment": false,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 131072, "output": 32768 }
        },
        "deepseek-v4-pro": {
          "name": "DeepSeek V4 Pro",
          "attachment": false,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 1048576, "output": 32768 }
        },
        "deepseek-v4-flash": {
          "name": "DeepSeek V4 Flash",
          "attachment": false,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 1048576, "output": 32768 }
        },
        "kimi-k2.7-code": {
          "name": "Kimi K2.7 Code",
          "attachment": true,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 262144, "output": 32768 }
        },
        "kimi-k2.6": {
          "name": "Kimi K2.6",
          "attachment": true,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 262144, "output": 32768 }
        },
        "qwen3.8-27b": {
          "name": "Qwen 3.8 27B",
          "attachment": true,
          "reasoning": true,
          "tool_call": true,
          "temperature": true,
          "limit": { "context": 262144, "output": 32768 }
        }
      }
    }
  }
}

Add your key through the API-key field of the provider form or opencode auth — it lives in opencode’s own auth store, not in this file. Hover a model in the picker to confirm: it should now show the real context limit and reasoning support.

Cline, Cursor, aider and friends

The same three values work everywhere; the only extra step worth taking is telling the tool the model’s context size and output budget wherever it lets you:

• Cline / Roo Code: choose the "OpenAI Compatible" provider, paste base URL + key + model ID, and fill the context-window and max-output fields from the table above.
• Cursor: Settings → Models → override the OpenAI base URL with the gateway’s, paste your key, and add the model IDs as custom models.
• aider: pass the base URL and key on the command line:

aider --openai-api-base https://api.clfaigateway.dev/v1 \
      --openai-api-key sk-gw-... \
      --model openai/glm-5.3

Three settings that prevent confusing failures

• Give the model room to answer after it thinks. Reasoning is spent from the same max_tokens budget as the answer, and on 100K+ prompts these models can think for thousands of tokens first. If your tool sets max_tokens itself, make it 8,000 or more for agentic work; if it sends nothing, the gateway defaults to 32,768 and you are fine. Details on the streaming page.
• A long first-token wait is not a hang. A 250K-token prompt spends ~20 seconds in prefill before anything streams, then the model may think for minutes before visible text. Keep client read timeouts at 60 s or more and turn on reasoning display if your tool has it — in our logs the most common "hang" is a client giving up mid-think.
• Only four models read images. Attach images only on the models marked Yes above — the others return a named 400 vision_not_available before you are charged anything. The full comparison is on the vision page.