Support per-agent lean local-model configuration
TL;DR
Added support for the experimental localModelLean configuration option on individual agents. This allows lean local-model mode to be enabled selectively rather than globally.
What changed
OpenClaw added support for the experimental localModelLean option on individual agents. Users can now set this flag in the YAML configuration for specific agents instead of applying it across the entire instance. The change rolled out on 22 May 2026 and targets self-hosted setups that mix local and remote models.
The option reduces memory and compute demands when an agent runs lighter local models. It keeps the global configuration untouched so other agents retain full model settings.
Why it matters
Selective lean mode gives Vibe Builders finer control over resource use on a single VPS. One agent can handle quick local tasks while another routes heavy work to paid APIs without forcing a full restart or config rewrite.
This move pressures global-only tools that lack per-agent switches. It bets that operators will run mixed workloads and prefer to tune costs at the agent level rather than the instance level.
How to use it
Edit the agent YAML file and add localModelLean: true under the target agent block. Restart only that agent through the CLI command openclaw agent restart .
The feature is available now on the latest main branch. No paid plan is required. Test the setting first on a non-critical agent to confirm token and memory savings before moving production tasks.
Watch for
Confirm the bet if per-agent lean mode cuts monthly VPS bills by 20 percent or more in mixed setups. It breaks if localModelLean introduces inconsistent context recall across agents. Expect a follow-up release that adds per-agent model routing rules next quarter.
Who this matters for
- Vibe Builders: Toggle localModelLean in agent YAML to cut VPS costs while keeping high-end APIs for core tasks.
- Developers: Implement per-agent configuration blocks to mix local and remote model resource demands dynamically.
Harsh’s take
Granular resource management is the only way to scale self-hosted agent swarms without ballooning infrastructure costs. OpenClaw moving away from global-only configurations shows a shift toward production-grade operations where every megabyte of VRAM matters. By allowing lean mode on a per-agent basis, builders can finally run 'utility' agents on cheap local hardware while reserving expensive compute for the heavy lifting.
This update highlights a growing gap between monolithic AI wrappers and modular frameworks. If your orchestration layer forces a global setting for all agents, it is effectively a toy. Operators need this level of precision to optimize latency and spend.
Expect more frameworks to follow suit as the industry moves toward heterogeneous model environments.
by Harsh Desai
About OpenClaw
View the full OpenClaw page →All OpenClaw updatesMore from OpenClaw
- App UpdateAdd doctor warnings for hidden MCP server tools
The doctor utility now warns users when sandbox tool policies hide configured Model Context Protocol (MCP) server tools before provider requests are made.
- App UpdateExpose stale-running task maintenance details in CLI
The tasks maintenance CLI command now outputs detailed JSON explanations for retained and reconcile candidates, detailing backing-session, cron, CLI, and wedged-subagent states.
- FeatureIntroduce bundled Policy plugin for workspace conformance
Added a bundled Policy plugin that enables policy-backed channel conformance checks, doctor lint findings, and opt-in workspace repair to streamline configuration management.