opencode-plugboard

Appearance

Karakeep ​

Karakeep saves bookmarks, pages and PDFs. An LLM tags them and, if you turn it on, summarizes them.

Upstream: source, docs.

Plugboard options ​

Give Karakeep its own key and two aliases. vision must accept images, because INFERENCE_IMAGE_MODEL sends them. Docker needs a reachable address (Serving other machines).

jsonc
{
  "options": {
    "listen": { "host": "0.0.0.0", "port": 4010 },
    "keys": { "karakeep": "{env:PLUGBOARD_KEY_KARAKEEP}" },
    "models": ["opencode-go/glm-5.3-flash", "anthropic/claude-sonnet-4-6"],
    "aliases": {
      "tag": "opencode-go/glm-5.3-flash",
      "vision": "anthropic/claude-sonnet-4-6"
    },
    "apps": { "karakeep": { "allowedModels": ["tag", "vision"] } }
  }
}

Connect ​

Set these in the .env that Karakeep's compose file reads:

ini
# The plugboard app key, not an OpenAI key.
OPENAI_API_KEY=<the karakeep app key>
OPENAI_BASE_URL=http://host.docker.internal:4010/v1

# Aliases from plugboard's `aliases` option.
INFERENCE_TEXT_MODEL=tag
INFERENCE_IMAGE_MODEL=vision

# Plugboard serves no embeddings. Keep semantic search off.
SEMANTIC_SEARCH_ENABLED=false

OPENAI_API_KEY takes the plugboard key. SEMANTIC_SEARCH_ENABLED=false hides the semantic and hybrid search modes, which need embeddings.

On Linux, add the host name to the web service in docker-compose.override.yml:

yml
services:
  web:
    extra_hosts:
      # Linux needs this to resolve host.docker.internal. Docker Desktop and Colima already provide it.
      - "host.docker.internal:host-gateway"

The files are in examples/self-hosted/karakeep/.

What works ​

  • Tagging, with INFERENCE_TEXT_MODEL. It is on by default.
  • Image tagging, with INFERENCE_IMAGE_MODEL.
  • Summaries, with INFERENCE_ENABLE_AUTO_SUMMARIZATION=true.
  • Structured output. INFERENCE_OUTPUT_SCHEMA defaults to structured, and plugboard answers it through a tool.

Limits ​

  • Semantic search, hybrid search and related bookmarks need embeddings, which plugboard does not serve. Leave them off, or point EMBEDDING_OPENAI_BASE_URL at an embedding provider you run.
  • Keep tools on for this app. A tools: false policy drops the schema (Per-app policy). If a model fails on structured output, set INFERENCE_OUTPUT_SCHEMA=json.
  • INFERENCE_CONTEXT_LENGTH defaults to 2048 tokens, so long pages are cut before tagging. Raise it.
  • A large import sends many calls at once. Set limits.perApp.requestsPerMinute on the karakeep app if your provider rate-limits.