Lemonade Dynamic Model Access for OpenCode

Automatically loads the model list from one or more remote Lemonade servers into OpenCode, so models never need to be hard-coded in opencode.json.

How it works

• opencode.json declares one provider entry per Lemonade server (using the @ai-sdk/openai-compatible adapter) with the server's base URL and an empty models map.
• lemonade-models.js is an OpenCode plugin. Its config hook runs at OpenCode startup, fetches GET <baseURL>/models from every provider named lemonade or lemonade-*, and injects the downloaded models into the config — including context window size and tool-calling / reasoning / vision capabilities derived from the server's model labels.

Restart OpenCode to pick up models added to or removed from the server.

Quick start

Install against a single Lemonade server (default port is 13305, so --host=hal9k means hal9k:13305):

$ ./setup.sh --host=hal9k:13305
Setting up Lemonade dynamic model access for OpenCode...
Checking http://hal9k:13305/api/v1 ... ok, 11 models
Installed plugin to /home/joe/.config/opencode/plugins/lemonade-models.js
Backed up existing config to /home/joe/.config/opencode/opencode.json.bak.20260731221150
Wrote /home/joe/.config/opencode/opencode.json

Setup complete!

Then verify OpenCode sees the models:

$ opencode models | grep lemonade
lemonade/DeepSeek-Qwen3-8B-GGUF
lemonade/Devstral-Small-2507-GGUF
lemonade/Qwen3-Coder-30B-A3B-Instruct-GGUF
...

And run a prompt against one of them:

$ opencode run --model lemonade/Qwen3-Coder-30B-A3B-Instruct-GGUF "Reply with exactly one word: pong"
pong

Inside the OpenCode TUI, pick a model with /models — they appear under the provider name, e.g. Lemonade (hal9k).

Installation examples

# Default: localhost:13305
./setup.sh

# One remote server
./setup.sh --host=hal9k:13305

# Several servers at once: the first becomes provider "lemonade",
# the others "lemonade-<hostname>"
./setup.sh --host=hal9k:13305 --host=deepthought:13305

# With an API key (Lemonade does not require one by default)
./setup.sh --host=hal9k:13305 --api-key="sk-..."
export LEMONADE_API_KEY="sk-..."   # add to ~/.bashrc so OpenCode can see it

--host writes a fresh config containing exactly the listed hosts. The script installs the plugin to ~/.config/opencode/plugins/ and backs up any existing ~/.config/opencode/opencode.json first.

The generated config looks like this — note the empty models map, which the plugin fills in at every OpenCode startup:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "lemonade": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Lemonade (hal9k)",
      "options": {
        "baseURL": "http://hal9k:13305/api/v1",
        "apiKey": "{env:LEMONADE_API_KEY}"
      },
      "models": {}
    }
  },
  "permission": {
    "read": "allow",
    "edit": "ask",
    "bash": "ask"
  }
}

Adding and removing servers

--add-host and --delete-host edit the existing config in place — other providers and settings are left untouched. Both are repeatable, deletions are applied before additions, and the previous config is backed up before every change.

$ ./setup.sh --add-host=deepthought
Checking http://deepthought:13305/api/v1 ... ok, 4 models
...
Added provider "lemonade-deepthought" (deepthought:13305)

$ ./setup.sh --delete-host=deepthought
Deleted provider "lemonade-deepthought" (deepthought:13305)

Hosts are matched by hostname:port in the provider's baseURL, so the same value used to add a host removes it — regardless of what the provider entry ended up being named. Adding a host that is already configured is a no-op:

$ ./setup.sh --add-host=hal9k
Provider "lemonade" already uses hal9k:13305; skipping

An unreachable host is added anyway (the plugin retries at every OpenCode startup), so you can configure servers that are currently down:

$ ./setup.sh --add-host=deepthought:13306
Checking http://deepthought:13306/api/v1 ... UNREACHABLE (continuing; the plugin will retry at OpenCode startup)
Added provider "lemonade-deepthought" (deepthought:13306)

Restart OpenCode after any change to refresh the model list.

Troubleshooting

Check what the server itself reports — this is exactly what the plugin reads:

$ curl -s http://hal9k:13305/api/v1/models | python3 -c \
    "import json,sys; [print(m['id']) for m in json.load(sys.stdin)['data']]"
DeepSeek-Qwen3-8B-GGUF
Devstral-Small-2507-GGUF
...

If models are missing from OpenCode but present in the curl output, run OpenCode with debug logging and look for lemonade-models: errors printed by the plugin (opencode --print-logs).

Uninstall

$ ./uninstall.sh
Removed Lemonade models plugin
Restored previous OpenCode configuration from opencode.json.bak.20260731221150
Uninstallation complete!

Removes the plugin and restores the most recent configuration backup made by setup.sh (or removes the generated config if no backup exists).