Ollama

The adapter-ollama adapter adds service-style Lua access to a local Ollama instance. It does not expose any HomeCmdr devices — it is accessed entirely through ctx:invoke(...) from scenes and automations.

Installation

# From your homecmdr-api workspace root
homecmdr pull adapter-ollama
cargo build

Then add to crates/adapters/src/lib.rs:

use adapter_ollama as _;

Configuration

[adapters.ollama]
enabled = true
base_url = "http://127.0.0.1:11434"
model = "llava"

Invoke Targets

All targets are accessed via ctx:invoke(target, payload) in Lua.

ollama:generate

Text generation.

local result = ctx:invoke("ollama:generate", {
  prompt = "Write a one-sentence weather summary.",
  model = "llama3",  -- optional, overrides config default
})
local text = result.response

Required: prompt. Optional: model, suffix, system, template, format, options, keep_alive, raw, images.

Response fields: response, boolean, done, and timing fields.

ollama:vision

Vision reasoning with an image.

local result = ctx:invoke("ollama:vision", {
  prompt = "Reply only true or false. Are clothes on the clothesline?",
  image_base64 = snapshot_base64,
})

if result.boolean == true then
  ctx:command("elgato_lights:light:0", { capability = "power", action = "on" })
end

Required: prompt, image_base64. Optional: model, system, template, format, options, keep_alive, raw.

Response fields: response, boolean, done, and timing fields.

ollama:chat

Multi-turn chat.

local result = ctx:invoke("ollama:chat", {
  messages = {
    { role = "system", content = "Be concise." },
    { role = "user",   content = "Summarize the weather in one sentence." },
  },
})
local reply = result.message.content

Required: messages. Optional: model, format, options, keep_alive, tools.

messages is a Lua list of message tables with fields: role, content, images, tool_calls, tool_name.

Response fields: message, done, and timing fields.

ollama:embeddings

Generate embeddings for one or more strings.

local result = ctx:invoke("ollama:embeddings", {
  input = {
    "front door open",
    "back door open",
  },
})
-- result.embeddings is a list of numeric arrays

Required: input (string or list of strings). Optional: model, keep_alive, options, truncate.

Response: embeddings (nested numeric arrays).

ollama:tags

List available models.

local result = ctx:invoke("ollama:tags", {})
-- result.models is an array of model info tables

No payload required.

ollama:ps

List currently loaded models.

local result = ctx:invoke("ollama:ps", {})

No payload required.

ollama:show

Show details for a model.

local result = ctx:invoke("ollama:show", {
  model = "llava",
  verbose = false,
})

Optional: model, verbose.

Returns the raw JSON object converted into Lua tables and lists.

ollama:version

Return the Ollama server version.

local result = ctx:invoke("ollama:version", {})
local version = result.version

No payload required.

Shared Helper Pattern

For repeated vision checks, a shared script module avoids duplication:

-- config/scripts/ollama.lua
local M = {}

function M.vision_bool(ctx, prompt, image_base64)
  local result = ctx:invoke("ollama:vision", {
    prompt = prompt,
    image_base64 = image_base64,
  })
  return result.boolean == true
end

return M

Usage in an automation:

local ollama = require("ollama")

if ollama.vision_bool(ctx, "Reply only true or false. Is the garage door open?", snapshot) then
  ctx:log("warn", "garage door left open")
end