OpenClaw on a local Lemonade server

OpenClaw is a personal AI assistant that can browse the web, manage files, run code, and orchestrate multi-step tasks on your behalf. By default it connects to cloud-hosted AI APIs, but you can run it entirely on your own hardware using Lemonade Server as the local backend, no API keys, no cloud costs, no data leaving your machine.

This script automates the full end-to-end setup: it installs Lemonade Server from the official PPA, pulls an AI model, installs OpenClaw, and wires the two together so OpenClaw uses your local Lemonade instance as its inference backend.

---

Prerequisites

• Ubuntu 24.04 (x86-64)
• Internet connection (for downloading packages and model weights)

---

Quick Start

curl -fsSL https://raw.githubusercontent.com/meghsat/lemonclaw_bash/main/openclaw-lemonade-setup.sh | bash

Environment overrides

All settings can be overridden via environment variables before running the script:

Variable	Default	Description
LEMONADE_BASE_URL	http://127.0.0.1:13305	URL the Lemonade server listens on
LEMONADE_PORT	13305	Lemonade server port
LEMONADE_MIN_CTX_SIZE	32768	Minimum context window size configured on the server
OPENCLAW_LEMONADE_MODEL_ID	(interactive)	Skip the model picker and use this model directly
OPENCLAW_LEMONADE_CONTEXT_TOKENS	190000	Context window size passed to OpenClaw
OPENCLAW_LEMONADE_MODEL_MAX_TOKENS	64000	Max output tokens per request
OPENCLAW_LEMONADE_MAX_AGENTS	2	Max concurrent top-level agents
OPENCLAW_LEMONADE_MAX_SUBAGENTS	2	Max concurrent subagents per agent
OPENCLAW_LEMONADE_GATEWAY_PORT	18789	Port the OpenClaw gateway listens on
OPENCLAW_LEMONADE_GATEWAY_BIND	loopback	Network interface for the gateway (loopback or all)
OPENCLAW_LEMONADE_SKIP_TUNING	0	Set to 1 to skip post-onboard config tuning

---

What the Script Does

Step 1 — Install Lemonade via PPA

Installs prerequisites (software-properties-common, curl, python3, wget), then adds the ppa:lemonade-team/bleeding-edge repository and installs lemonade-server via apt. Skipped if lemonade is already present on PATH.

Step 2 — Start Lemonade Server

Pings http://127.0.0.1:13305/api/v1/models to check whether the server is already running. If not, launches lemond (or lemonade-server as a fallback) in the background via nohup, then polls until the endpoint responds — up to 10 attempts, 2 seconds apart. Logs go to /tmp/lemonade-server.log.

Step 2b — Configure context size

Reads the current ctx_size from lemonade config and, if it is below LEMONADE_MIN_CTX_SIZE (default 32768), runs lemonade config set ctx_size=<value> to raise it.

Step 3 — Model selection

The script first recommends the Qwen3.5-35B-A3B-Q4_K_M model (~22 GB) with an interactive Yes/No prompt:

• Yes: use the recommended model: runs lemonade import with a JSON spec that registers the model (checkpoint, recipe, labels, size, and ctx_size=32768), then lemonade pull to download it.
• No: choose a different model: displays the full lemonade list output and shows an interactive arrow-key menu of all available models and their download status.

• If only one model is available it is selected automatically.
• If the chosen model has not been downloaded yet, you are offered the option to lemonade pull it before continuing (this can take a while for large GGUF files).
• If switching to a different model than what was previously configured, the old model is unloaded first to free VRAM.
• The model can be pre-selected non-interactively via OPENCLAW_LEMONADE_MODEL_ID.

Step 4 — Install OpenClaw

Sets ~/.npm-global as the npm global prefix (so the binary lands in a user-writable location), then runs the official OpenClaw installer:

curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-prompt --no-onboard

Skipped if openclaw is already on PATH.

Step 5 — Configure OpenClaw (non-interactive onboard)

Runs openclaw onboard --non-interactive to point OpenClaw at the local Lemonade server:

Onboard flag	Value
--mode	local
--custom-base-url	http://127.0.0.1:13305/api/v1
--custom-provider-id	lemonade
--custom-compatibility	openai (Lemonade exposes an OpenAI-compatible API)
--custom-api-key	lemonade (placeholder; no real key needed)
--custom-model-id	the model selected in Step 3
--gateway-port	18789
--gateway-bind	loopback

Before running, the script backs up any existing ~/.openclaw/openclaw.json to a timestamped .bak file. The step is skipped if the config already contains a lemonade provider entry for the selected model.

Step 6 — Auto-tune configuration

Edits ~/.openclaw/openclaw.json with Python to apply model-specific settings that the non-interactive onboard does not set:

• Sets the agent default model to lemonade/<model-id>
• Sets contextTokens to 190000
• Sets agent and subagent maxConcurrent values
• Sets contextWindow, maxTokens, zero-cost pricing, and reasoning: true on the model entry
• Configures memory-search embeddings to use nomic-embed-text-v1-GGUF via Lemonade's OpenAI-compatible /api/v1/embeddings endpoint (provider: openai)
• Sets the default browser profile to connect to Chrome via CDP on http://127.0.0.1:9222

Step 6b — Install Google Chrome

OpenClaw can control a browser for web tasks. If no Chrome or Chromium binary is found, the script adds the Google Chrome APT repository and installs google-chrome-stable.

Step 7 — Interactive onboard and hatch

Launches openclaw onboard --auth-choice skip interactively (reading from /dev/tty) so you can configure:

• The OpenClaw gateway (auth token, allowed origins)
• Hooks (shell commands that fire on agent events)
• Skills (custom agent capabilities)
• Notification channels

After that completes:

• A TOOLS.md file is written to ~/.openclaw/workspace/ documenting how the agent can use Chrome (CDP on port 9222) and noting the Lemonade backend.
• The memory search index is built with openclaw memory index.
• Chrome is launched in the background with --remote-debugging-port=9222 and the OpenClaw dashboard URL is opened automatically.
• A second pass of openclaw onboard runs the hatching flow (the initial agent interaction experience).

Step 8 — Start OpenClaw gateway (foreground)

Starts the OpenClaw gateway in the foreground:

openclaw gateway run --bind loopback --port 18789 --force

The gateway is the local HTTP server that the OpenClaw web UI and browser extension talk to.

---