OpenClaw March 2026: ClawHub, Sub-Agents & More

ClawHub: the agent marketplace

ClawHub is a curated marketplace of pre-built OpenClaw agents and tools. Instead of configuring an agent from scratch for email triage, calendar management, or code review, you pull a community-vetted agent from ClawHub and drop it into your workflow. Think of it as npm, but for agent behaviors.

The quality bar matters here. These aren't raw prompt templates — ClawHub agents are tested against real tasks, versioned, and community-rated. A ClawHub integration for Notion, Linear, or Slack has already been validated against the edge cases you'd otherwise discover on your own after two weeks of debugging.

Token implications: more capable, pre-built agents typically compose multiple tools and sub-tasks in sequence. A ClawHub research agent might run a web search, summarize results, cross-reference a second source, and format a report — all in one invocation. That's a longer task chain than a single-step prompt, which means higher token usage per task. The capability gain is real, but so is the cost. Routing lighter sub-tasks to smaller models becomes more important as you add ClawHub integrations.

Multi-model sub-agents

OpenClaw can now delegate sub-tasks to different models. A primary Claude Sonnet agent can spawn Haiku sub-agents for lightweight operations — formatting, classification, short summarization — and reserve Sonnet for the work that actually requires deeper reasoning. The architecture is explicit: you define which model handles which class of sub-task.

This is smart routing at the agent architecture level, and it closes a gap that's been expensive for teams running complex workflows. Previously, every sub-task inherited the cost of the primary model. A Sonnet agent doing data extraction, field mapping, and report formatting paid Sonnet prices for all three steps — even though the first two don't need that capability.

What this means in practice: a well-configured multi-model setup can cut per-task token spend by 40–60% on workflows with a mix of heavy and light sub-tasks. The catch is that configuring model routing manually adds complexity. Rapid Claw's smart routing layer handles this automatically — the same principle, applied across your entire agent fleet without per-agent configuration.

Adjustable sub-agent thinking budgets

You can now set the reasoning budget for each sub-agent independently. More thinking tokens allocated to a sub-agent means better results on ambiguous or multi-step tasks — and a higher token bill. Less thinking means faster, cheaper responses for straightforward sub-tasks where extended reasoning adds no value.

This control is important. Before this update, sub-agent reasoning was opaque — you couldn't tune it without forking the agent code. Now you can set a budget per task type and observe the quality-cost tradeoff directly. A code review sub-agent might warrant a high thinking budget. A date-parsing sub-agent does not.

The discipline required here is the same as any resource budget: you need to make the decision deliberately rather than leaving it at the default. Most teams will over-allocate thinking budget initially and adjust down after reviewing their token logs. Start conservative and increase only where output quality degrades.

The /btw command

/btw opens a side conversation with your agent without interrupting the main task flow. If your agent is mid-way through a multi-step research task and you want to add context or redirect part of the scope, /btw surfaces that input without canceling or restarting the current task.

The practical value is clearer in longer-running tasks. A 20-step workflow doesn't need to restart from step one because you realized at step 12 that the output format should change. /btw lets the agent absorb the update and continue. For teams running automated overnight jobs, it also enables lightweight human-in-the-loop checkpoints without designing full approval gates into the workflow.

Expanded platform support

This update adds native integrations for Slack, Microsoft Teams, Discord, and WhatsApp. OpenClaw agents can now receive tasks, send updates, and surface results directly inside the platforms your team already uses — without routing through a separate dashboard. OpenAI Gateway compatibility issues present in the previous release have also been resolved.

One milestone worth noting: OpenClaw hit 100,000 GitHub stars in February, overtaking several well-established open-source projects in under a year. The ecosystem momentum has material implications — more integrations, faster bug resolution, and a broader pool of ClawHub agents from contributors.

What this means for hosting costs

Each of these features increases the potential token consumption of an OpenClaw deployment. More ClawHub integrations means more complex task chains. Multi-model sub-agents means more parallel model calls. Expanded platform connections means more entry points for tasks to originate. If you're on a pay-per-token model with no routing layer, this update will show up in your bill.

Rapid Claw's $29/mo plan includes $20 in AI tokens with smart routing enabled by default. When your token usage exceeds the included amount, the system throttles and notifies you rather than running up an unexpected overage charge. As agent complexity grows with features like ClawHub and multi-model routing, that predictability matters more, not less.

Smart routing becomes specifically more valuable as sub-agent count grows. A single-agent workflow with one model is easy to cost-project. A ClawHub workflow spawning four sub-agents across three models with variable thinking budgets is not — unless a routing layer is normalizing the decisions automatically. That's the case Rapid Claw was built for.