Skip to content

Lemonade Runner

Lemonade is kew's own tool-use loop (kew-lemonade) over any OpenAI-compatible local endpoint. It supports AMD Lemonade, LM Studio, llama.cpp's llama-server, vLLM, Ollama, and any other OpenAI-compatible server. Cost is $0 (local); nothing leaves the machine.

Prerequisites

  1. Install a local OpenAI-compatible server — see Lemonade's installation guide.
  2. Load a model into your server.

Selecting a model

Set model = "lemonade/<model-id>" in kew.toml; the lemonade/ prefix routes to this runner.

[agent]
model = "lemonade/llama3"

Configuration

The [agent.lemonade] block accepts the following settings:

  • base_url (default http://localhost:13305/v1) — the OpenAI-compatible endpoint to connect to.
  • max_steps (default 10) — the maximum number of tool calls allowed in a single run.

Examples

A kew.toml snippet:

[agent]
model = "lemonade/llama3"

[agent.lemonade]
base_url = "http://localhost:13305/v1"
max_steps = 15

Direct invocation with kew-lemonade:

kew-lemonade --base-url http://localhost:13305/v1 --model llama3 -p "Write a Python script to calculate Fibonacci numbers"

Limitations (v1)

  • Model quality is the model's — kew does not influence model quality; it's entirely up to the model you choose.
  • Tool-calling reliability varies by model — some models may struggle with consistent tool calling.
  • run_bash runs with the loop's privileges — since this runner executes locally, run_bash commands run with the same privileges as the kew process, so it's recommended to run kew inside a sandbox.