OpenCode with Plan2API: custom provider setup

Settings at a glance

Here are the values you need for the provider block in opencode.json. OpenCode is MIT-licensed and accepts any OpenAI-compatible API as a custom provider.

Set up OpenCode with Plan2API

1. Sign in with a Google account at plan2api.com/connect, connect your Claude Pro or Max, ChatGPT, Google AI Pro or Ultra, or Kimi Code plan, and generate a p2a_ API key.
2. Install OpenCode with curl -fsSL https://opencode.ai/v2/install | bash, npm i -g @opencode/cli@latest, or brew install anomalyco/tap/opencode-v2.
3. Export your key as PLAN2API_API_KEY in the shell where you run OpenCode, or if you'd rather not use an environment variable, run /connect inside OpenCode, choose Other, set the provider id to plan2api, and paste the key so OpenCode saves it to ~/.local/share/opencode/auth.json.
4. Save the opencode.json file from the next section in your project root (which overrides your global config) or globally at ~/.config/opencode/opencode.json.
5. Start opencode and run /models to select a Plan2API model.

# Install OpenCode
curl -fsSL https://opencode.ai/v2/install | bash

# Make your Plan2API key available to OpenCode
export PLAN2API_API_KEY="p2a_your_key"

# Start OpenCode in your project
opencode

The opencode.json config

Drop this opencode.json provider config into your project root or ~/.config/opencode/opencode.json to wire up your models and set your defaults.

• The provider id plan2api becomes the prefix of every model id you call in OpenCode.
• Each key under models must match the upstream Plan2API model id, including its provider prefix like claude/ or gemini/.
• Setting small_model to plan2api/gemini/gemini-3.8-flash-high routes light background tasks to Gemini 3.8 Flash so they don't eat into your Claude quota.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "plan2api": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Plan2API",
      "options": {
        "baseURL": "https://api.plan2api.com/v1",
        "apiKey": "{env:PLAN2API_API_KEY}"
      },
      "models": {
        "claude/claude-opus-5-5": { "name": "Claude Opus 5.5" },
        "openai/gpt-6.1-sol": { "name": "GPT-6.1 Sol" },
        "gemini/gemini-4-argon-high": { "name": "Gemini 4 Argon" },
        "gemini/gemini-3.8-flash-high": { "name": "Gemini 3.8 Flash" },
        "kimi/k3": { "name": "Kimi K3" }
      }
    }
  },
  "model": "plan2api/claude/claude-opus-5-5",
  "small_model": "plan2api/gemini/gemini-3.8-flash-high"
}

Pick your models

You can mix models from Claude, ChatGPT Codex, Gemini and Kimi under the same p2a_ key, and the table below lists the main ids and token limits. OpenCode reads the part before the first slash as the provider id and keeps everything after it as the model id, so the double slash in plan2api/claude/claude-opus-5-5 is right. Run GET /v1/models with your key to see every model id Plan2API serves, including the openai/<model> Codex entries; models from providers you haven't connected return HTTP 403 no_credentials.

What works and what doesn't

• Streaming, tool and function calling, and reasoning_effort (or reasoning.effort) work across Claude, Codex, Gemini and Kimi models.
• Image input works on Claude, Gemini and Kimi models, plus Codex models that accept images, with up to 32 image URLs fetched per request (up to 20 MB and 15 seconds each).
• Set npm to @ai-sdk/openai-compatible rather than @ai-sdk/openai, because @ai-sdk/openai calls /v1/responses, which Plan2API doesn't serve.
• Plan2API does not serve /v1/embeddings, /v1/completions, or Anthropic's /v1/messages.
• JSON mode (response_format set to json_object or json_schema) works on Codex, Kimi and Gemini, but returns HTTP 400 unsupported_parameter on Claude models.
• temperature and top_p are ignored on Claude and Codex models, passed through on Gemini, and fixed at 1 and 0.95 on Kimi.
• The Free plan includes 50 chat requests per day (resetting at 00:00 UTC), and every tool-call round trip OpenCode makes counts as one request.

Troubleshooting

When a request fails, Plan2API returns an OpenAI-style error object with message, type and code (or a final server-sent event if streaming has already started), and the table below covers the common status codes.