OpenCode Custom Providers: Connect DeepSeek, GLM, Kimi, Qwen

OpenCode is one of the most popular open-source coding agents — a terminal-first tool, provider-agnostic by design. Through the AI SDK and Models.dev it supports dozens of providers out of the box, and for anything it doesn’t list, you can wire up a custom provider in a few lines. That makes it a natural home for cheap models like DeepSeek, GLM, Kimi, and Qwen.

This covers the config that connects any OpenAI-compatible endpoint, where the files live, and the small things that go wrong. For why you’d run a cheaper backend at all, see use cheap AI models with Claude Code — the cost logic carries straight over.

Two ways to add a provider#

OpenCode supports many providers natively. For an OpenAI-compatible one, you have two paths:

• Quick: opencode auth login, scroll to Other, paste your key and base URL.
• Explicit: define a custom provider in the config — clearer, reproducible, and version-controllable.

The config approach is worth knowing because it pins exact model IDs and base URLs.

Install OpenCode#

npm install -g opencode-ai

On Windows it works directly, but WSL gives the cleaner shell — see install OpenCode on Windows with WSL and the WSL install guide.

Add a custom provider#

OpenCode reads a global config (commonly ~/.config/opencode/opencode.json) and a project-level opencode.json. A custom DeepSeek provider using the OpenAI-compatible SDK looks like this:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "deepseek": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepSeek",
      "options": {
        "baseURL": "https://api.deepseek.com/v1",
        "apiKey": "{env:DEEPSEEK_API_KEY}"
      },
      "models": {
        "deepseek-chat": { "name": "DeepSeek Chat" },
        "deepseek-reasoner": { "name": "DeepSeek Reasoner" }
      }
    }
  }
}

The {env:DEEPSEEK_API_KEY} syntax pulls the key from your environment instead of hard-coding it. Export it first:

export DEEPSEEK_API_KEY="sk-your-key"

Other providers#

Same structure, different endpoint. Qwen via Alibaba’s international OpenAI-compatible URL:

{
  "provider": {
    "dashscope": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Qwen",
      "options": {
        "baseURL": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
        "apiKey": "{env:DASHSCOPE_API_KEY}"
      },
      "models": { "qwen3-coder-plus": { "name": "Qwen3 Coder Plus" } }
    }
  }
}

For per-model specifics, see DeepSeek + OpenCode, GLM + OpenCode, Kimi K2.5 Turbo + OpenCode, and MiniMax M2 + OpenCode.

Pick the model and run#

Set a default model in the config or choose it in the TUI:

{
  "model": "deepseek/deepseek-chat"
}

Then launch:

opencode

The model identifier is always provider/model, matching your config keys.

Troubleshooting#

• Provider not found — the provider key and the model prefix must match exactly (deepseek ↔ deepseek/...).
• Auth failures — the {env:...} variable isn’t set in the shell that launched OpenCode.
• Wrong base URL — most providers expect the /v1 suffix; confirm on their docs.
• Model ID rejected — pull the current model name from the provider; IDs change between releases.

Wrapping up#

OpenCode runs on any cheap model through one config pattern: a custom provider using @ai-sdk/openai-compatible, with the base URL, an {env:...} API key, and your model list. Set a default provider/model and launch. The same block works for DeepSeek, Qwen, GLM, Kimi, and local servers — only the URL and IDs change.

For the bigger landscape, see best cheap Claude Code alternatives for Windows and the head-to-head in Claude Code vs Codex vs OpenCode vs Aider.