Add experimental local model lean mode
TL;DR
Introduce `agents.defaults.experimental.localModelLean: true` to disable heavyweight tools like browser and cron, reducing prompt size for weaker local models.
## What changed OpenClaw added an experimental local model lean mode on 18 May 2026. The new flag agents.defaults.experimental.localModelLean set to true disables heavyweight tools such as browser control and cron jobs.
This reduces overall prompt size so weaker local models can run without hitting context limits. The change appears in the latest CLI release and requires only a YAML edit plus restart.
## Why it matters The update makes self-hosted OpenClaw practical for users who prefer local models over paid API calls. It removes the constant pressure to upgrade hardware or accept slower responses on modest setups.
The move bets that many Vibe Builders will keep the agent running continuously once token bloat is under control. It directly challenges cloud-only agents that assume unlimited context and always-on tool access.
## How to use it Open the agents.yaml file in your OpenClaw install directory. Add agents.defaults.experimental.localModelLean: true under the defaults section, then restart the service.
The flag is live in the current GitHub main branch and works on any supported local model provider. Start with a single test workflow to confirm tool access matches expectations.
## Watch for Confirm the bet if prompt token counts drop by at least 30 percent in logs while core chat and memory functions stay intact. The setup breaks if remaining tools still overflow smaller models. The next logical step is granular per-tool toggles so users can re-enable browser only when needed.
Harsh’s take
For a solo Vibe Builder this flag simply removes the parts of OpenClaw that were already too heavy for local hardware. You lose browser automation and scheduled tasks, which means you now split those jobs across separate scripts or a second lighter agent.
The honest trade-off is speed and lower RAM use against reduced autonomy. Most day-to-day memory and chat work survives, but any workflow that relied on the disabled tools needs manual replacement.
Do this now: add the lean flag to your config, run a week of normal conversations, and measure both response latency and actual token spend before deciding whether to keep full tools on a paid endpoint.
by Harsh Desai
About OpenClaw
View the full OpenClaw page →All OpenClaw updatesMore from OpenClaw
- FeatureExpand QA-Lab with runtime parity scenarios
Added comprehensive runtime parity tiers and token-efficiency artifacts to the QA-Lab, including specific checks for Codex-vs-Pi compatibility and tool fixture coverage.
- App UpdateUpdate Node.js requirement and Pi packages
Raised the minimum supported Node.js version to 22.19 and updated Pi packages to version 0.75.1 to ensure compatibility with the latest runtime features.
- App UpdateOptimize Gateway startup and restart latency
Reduced restart ready latency by overlapping startup logging and plugin-service initialization with channel sidecars while maintaining strict readiness gating.