> ## Documentation Index
> Fetch the complete documentation index at: https://docs.powertokens.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# opencode setup guide

> Connect opencode to the powertokens platform through an OpenAI-compatible custom provider.

opencode manages models through `provider` configuration. When connecting to the `powertokens platform`, start with an OpenAI-compatible custom provider that points opencode at the platform `/v1/chat/completions` compatible endpoint.

This page is for opencode text-agent workflows. Some image, speech, and video task models in `/v1/models` also declare `openai`, but that does not mean they are text Chat Completions models for opencode.

If you need to choose the endpoint first, see [Text model protocols and endpoints](/en/ecosystem-tools/text-model-protocols).

## Choose a path

| Goal                  | opencode provider                             | Base URL                        | Recommended model                                 |
| --------------------- | --------------------------------------------- | ------------------------------- | ------------------------------------------------- |
| Recommended text path | custom provider + `@ai-sdk/openai-compatible` | `https://api.powertokens.ai/v1` | `glm-5.2`                                         |
| MiniMax text path     | custom provider + `@ai-sdk/openai-compatible` | `https://api.powertokens.ai/v1` | `MiniMax-M3`                                      |
| Claude/Anthropic path | `@ai-sdk/anthropic` custom provider           | `https://api.powertokens.ai/v1` | Use only when you need to validate `/v1/messages` |

## Prepare the API key

Store the `powertokens platform` API key in an environment variable:

```bash theme={null}
export POWERTOKENS_API_KEY="YOUR_POWERTOKENS_API_KEY"
```

You can also run `/connect` in the opencode TUI and save credentials there. The provider ID in `opencode.json` must match the provider ID used by `/connect`.

## OpenAI-compatible provider example

Create or update `.opencode/opencode.json` in your project:

```json theme={null}
{
  "$schema": "https://opencode.ai/config.json",
  "model": "powertokens/glm-5.2",
  "small_model": "powertokens/glm-4.5-air",
  "provider": {
    "powertokens": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "powertokens",
      "options": {
        "baseURL": "https://api.powertokens.ai/v1",
        "apiKey": "{env:POWERTOKENS_API_KEY}"
      },
      "models": {
        "glm-4.5-air": {
          "name": "glm-4.5-air"
        },
        "glm-5.2": {
          "name": "glm-5.2"
        },
        "MiniMax-M3": {
          "name": "MiniMax-M3"
        },
        "deepseek-v3-2-251201": {
          "name": "deepseek-v3-2-251201"
        }
      }
    }
  }
}
```

This path uses the platform `POST /v1/chat/completions` endpoint.

## Model choice

Use `glm-5.2` as the OpenAI-compatible default. Use `MiniMax-M3` for MiniMax scenarios. When you need BytePlus or Seed text models, prefer the OpenAI-compatible provider.

Do not add `image-generation`, `openai-video`, TTS, Kling, Vidu, or Wan task models to an opencode text provider. For the full text model matrix, see [Text model protocols and endpoints](/en/ecosystem-tools/text-model-protocols).

## Anthropic provider example

For Claude/Anthropic Messages, configure a separate `@ai-sdk/anthropic` custom provider. This path uses the platform `/v1/messages` endpoint; validate the current opencode version and target model end to end before relying on it.

```json theme={null}
{
  "$schema": "https://opencode.ai/config.json",
  "model": "powertokens-anthropic/glm-5.2",
  "small_model": "powertokens-anthropic/glm-5-turbo",
  "provider": {
    "powertokens-anthropic": {
      "npm": "@ai-sdk/anthropic",
      "name": "powertokens Anthropic",
      "options": {
        "baseURL": "https://api.powertokens.ai/v1",
        "apiKey": "{env:POWERTOKENS_API_KEY}"
      },
      "models": {
        "glm-5.2": {
          "name": "glm-5.2"
        },
        "glm-5-turbo": {
          "name": "glm-5-turbo"
        },
        "MiniMax-M3": {
          "name": "MiniMax-M3"
        },
        "deepseek-v3-2-251201": {
          "name": "deepseek-v3-2-251201"
        }
      }
    }
  }
}
```

Use `glm-5.2` as the default Anthropic-path model. Use `MiniMax-M3` for MiniMax-path validation. `deepseek-v3-2-251201` passed opencode Anthropic-provider tool-workflow validation. Seed models except `seed-1-6-flash-250715` also passed re-validation and can be selected for real-task validation. Do not use `seed-1-6-flash-250715` as an opencode Anthropic-provider default.

## Verify

First confirm opencode can read the provider and models:

```bash theme={null}
opencode models powertokens
opencode models powertokens-anthropic
```

Then verify the OpenAI-compatible path in non-interactive mode:

```bash theme={null}
opencode run --model powertokens/glm-5.2 "Reply exactly: ok"
```

You can also validate the same OpenAI-compatible path with a MiniMax text model:

```bash theme={null}
opencode run --model powertokens/MiniMax-M3 "Reply exactly: ok"
```

Start opencode:

```bash theme={null}
opencode
```

In the TUI, run:

```text theme={null}
/models
```

Select `powertokens/glm-5.2` or `powertokens-anthropic/glm-5.2`, then send:

```text theme={null}
Reply with ok and print the active model name.
```

## Capability boundaries

* The opencode `@ai-sdk/openai-compatible` provider covers text Chat Completions semantics only.
* A model declaring `openai` in `supported_endpoint_types` is not automatically suitable for opencode text-agent use; exclude image, speech, video, and task models.
* Reasoning models such as `MiniMax-M3` may return thinking content or consume extra output tokens. Validate the display behavior with real tasks.
* A model declaring `anthropic` in `supported_endpoint_types` is not automatically suitable for the opencode Anthropic provider. opencode sends tool metadata and uses streaming output, so validate the full tool workflow.
* `deepseek-v3-2-251201` passed an opencode Anthropic-provider task with tool use.
* `seed-1-6-flash-250715` is not supported by its upstream provider on `/v1/messages`; prefer the OpenAI-compatible provider.
* Do not write a real API key into `.opencode/opencode.json` or commit it to your repository.