<!-- Lazu · OpenCode custom provider: Claude, GPT and DeepSeek on Lazu · https://lazu.ai/blog/opencode-custom-provider -->

# OpenCode custom provider: Claude, GPT and DeepSeek on Lazu

2026-11-18 · Lazu team

Short answer: add a `provider.lazu` block to `~/.config/opencode/opencode.json` with `baseURL` set to `https://api.lazu.ai/v1`. One Lazu key then lets OpenCode switch across Claude, GPT, Gemini, DeepSeek, Kimi, GLM, and Qwen, while keeping a separate lightweight `small_model` for background tasks.

## The quick way: one script

In [Console › API Keys](https://lazu.ai/console/token), pick a key, click Connect, select OpenCode, and copy the one-line setup command (or run this directly):

```bash
curl -fsSL https://api.lazu.ai/setup/opencode | sh -s -- \
  --model claude-sonnet-5-5 \
  --light claude-haiku-5-5 \
  --also gpt-6.1-sol \
  --also deepseek-v4-pro
```

The script prompts for your key interactively (keeping it out of shell history), merges the `lazu` provider into `~/.config/opencode/opencode.json`, and runs a quick connectivity check. A PowerShell command is in the same modal on Windows.

## Manual setup in `opencode.json`

If you prefer keeping secrets in environment variables so your config file stays clean, export `LAZU_API_KEY` in your shell profile (`~/.zshrc` or `~/.bashrc`):

```bash
export LAZU_API_KEY="sk-lazu-..."
```

Then edit `~/.config/opencode/opencode.json` (or `$XDG_CONFIG_HOME/opencode/opencode.json` if `XDG_CONFIG_HOME` is set):

```json
{
  "$schema": "https://opencode.ai/config.json",
  "model": "lazu/claude-sonnet-5-5",
  "small_model": "lazu/claude-haiku-5-5",
  "provider": {
    "lazu": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Lazu",
      "options": {
        "baseURL": "https://api.lazu.ai/v1",
        "apiKey": "{env:LAZU_API_KEY}"
      },
      "models": {
        "claude-opus-5-5": { "name": "Claude Opus 5.5" },
        "claude-sonnet-5-5": { "name": "Claude Sonnet 5.5" },
        "claude-haiku-5-5": { "name": "Claude Haiku 5.5" },
        "gpt-6-luna": { "name": "GPT-6 Luna" },
        "deepseek-v4-pro": { "name": "DeepSeek V4 Pro" },
        "deepseek-flash": { "name": "DeepSeek Flash" },
        "kimi-k3": { "name": "Kimi K3" },
        "glm-5.3": { "name": "GLM-5.3" },
        "qwen3.8-max": { "name": "Qwen 3.8 Max" }
      }
    }
  }
}
```

Start `opencode` and run `/models` to switch between any listed model. Check the [model catalog](https://lazu.ai/models) for the full list of model IDs.

## Why split `model` and `small_model`

OpenCode sends frequent background requests for session titles and lightweight summaries. When `small_model` is unset, those calls inherit your primary `model` — meaning a session running `claude-opus-5-5` pays flagship rates just to title a thread.

Pointing `small_model` at `lazu/claude-haiku-5-5` or `lazu/deepseek-flash` keeps your main coding loop on a frontier model and routes background chores to a fast, low-cost model under the same key budget and usage log.

## Protocol notes for GPT-6 models

`@ai-sdk/openai-compatible` speaks `/v1/chat/completions`. Lazu's protocol layer bridges those requests losslessly to Anthropic's Messages API for Claude and to native endpoints for DeepSeek, Kimi, GLM, Qwen, and Gemini, with streaming and tool calls preserved (see [how Lazu translates APIs](/blog/lazu-protocol-layer)).

One upstream rule is worth knowing: OpenAI's reasoning-only GPT-6 variants (`gpt-6.1-sol`, `gpt-6-astra`) reject function tools on `/v1/chat/completions` and require `/v1/responses`. Under `@ai-sdk/openai-compatible`, pick `gpt-6-luna` or `gpt-6-sol` for tool-calling OpenAI sessions, or configure `@ai-sdk/openai` (Responses mode) / use [Codex](/blog/codex-custom-provider) or [Claude Code](/blog/claude-code-custom-api) when running `gpt-6.1-sol` with tools.

## Troubleshooting

- **Model missing from `/models`:** Verify the key under `provider.lazu.models` matches the exact ID in the [Lazu model catalog](https://lazu.ai/models) (case-sensitive, e.g. `MiniMax-M2.7`).
- **`401` authentication error:** If you use `"{env:LAZU_API_KEY}"`, verify `echo $LAZU_API_KEY` is non-empty in the shell where you launch `opencode`.
- **Sessions not grouped under OpenCode in the console:** Newer OpenCode releases changed their `User-Agent` format (`opencode/<channel>/<version>/cli`) and send `x-opencode-session-id`. Lazu updated its detection rules in `v0.7.260` ([details](/blog/coding-agent-user-agents)); updating OpenCode ensures every session is grouped cleanly.
