opencode-plugboard

Appearance

OpenCode ​

Another OpenCode (on your laptop, say) can use a plugboard hub as its provider. That OpenCode runs the tools on its own machine; the hub only answers the model calls.

Anthropic dialect (preferred) ​

The Messages route keeps reasoning and replays signed thinking with no extra setting. Point the built-in anthropic provider at the hub:

jsonc
{
  "$schema": "https://opencode.ai/config.json",
  "providers": {
    "anthropic": {
      "settings": { "baseURL": "http://127.0.0.1:4010/v1", "apiKey": "{env:PLUGBOARD_KEY}" },
      "models": {
        "fast": { "name": "fast", "limit": { "context": 200000, "output": 8000 } }
      }
    }
  }
}

OpenAI-compatible dialect ​

jsonc
{
  "$schema": "https://opencode.ai/config.json",
  "providers": {
    "plugboard": {
      "name": "plugboard",
      "package": "@ai-sdk/openai-compatible",
      "settings": { "baseURL": "http://127.0.0.1:4010/v1", "apiKey": "{env:PLUGBOARD_KEY}" },
      "models": {
        "fast": { "name": "fast", "limit": { "context": 200000, "output": 8000 } }
      }
    }
  }
}

This dialect drops reasoning unless the hub sets reasoningOutput: true (Reasoning output).

Providers go under providers (plural). Run with -m anthropic/fast or -m plugboard/fast; use any name GET /v1/models lists.

Limits and cost ​

A custom model without limit gets 200k context and 32k output, and without cost its spend shows as zero. Generate both from the hub:

sh
PLUGBOARD_KEY=<app key> bun .github/scripts/opencode-models.ts http://127.0.0.1:4010

It prints a models block with limit and cost for each model from /v1/models (run it from a plugboard checkout). Models with a #variant suffix are skipped: declare the base model.

Tools ​

OpenCode sends about 12 tools, with MCP tools behind the single Code Mode execute tool, so requests stay under maxTools (64). A server with codemode: false sends its tools directly and can go over; raise maxTools on the hub if requests are refused with Request has N tools; plugboard allows 64.

Prompt cache ​

Turn on the anchor on the hub so the provider's prompt cache keeps hitting across turns:

jsonc
"promptCache": { "affinity": "anchor", "scope": "app" }

See Prompt-cache affinity.

Scratch checks ​

opencode run --standalone -m plugboard/fast "reply with exactly: pong" uses your scratch config instead of the managed background service.