OpenClaw with Ollama Herd

TL;DR

• OpenClaw (formerly Clawdbot / Moltbot) turns a local model into an agent that reaches you through your messaging apps. Ollama ships first-party support via ollama launch openclaw.
• Install the Herd skill: clawhub install ollama-herd. Any OpenClaw-compatible agent can then manage your fleet in natural language.
• Point OpenClaw at the herd instead of one Ollama, and every agent request routes across all your Macs.
• The real win is context. OpenClaw recommends 64k+ context. Long agent sessions are exactly where local models fall over, and that is the problem Herd's three-layer context management was built for.

Start with Ollama's own path

If you have not run OpenClaw before, use Ollama's built-in launcher first. It is the shortest path to a working agent and it does not need Herd:

ollama launch openclaw

That configures OpenClaw against your local Ollama and gets you an agent on one machine. If one Mac is all you have, you are done, and this guide has nothing to add. Come back when a single machine starts being the bottleneck.

Install the Ollama Herd skill

Ollama Herd publishes a skill on ClawHub, OpenClaw's skill registry. Install it once and any OpenClaw-compatible agent gains fleet awareness:

clawhub install ollama-herd

Or find it by search:

clawhub search "ollama herd"

With the skill installed, your agent can answer questions like "which node is busiest right now", "pull qwen3-coder onto the Studio", or "why did that request fail", without you memorising endpoints. Full skill reference on the Agent Skill page.

Point OpenClaw at your herd

OpenClaw talks to an Ollama-compatible endpoint. Ollama Herd is one, so this is a URL change rather than an integration:

# Instead of a single Ollama on one machine:
#   http://localhost:11434
#
# Point at the herd:
    http://localhost:11435

Nothing else in your OpenClaw config changes. Requests now get scored across every registered Mac and land on whichever one can answer fastest. See Quickstart if the herd is not running yet.

Why a fleet matters for agents specifically

Agent workloads are different from chat. A chat turn is short and forgiving. An OpenClaw session runs for hours, accumulates a long history, and fires many tool calls, and every one of those is a full inference request.

• Long sessions are context-hungry. OpenClaw's own guidance recommends at least 64k context. A local model at 64k+ is slow to prefill and quick to wedge, which is precisely the failure mode Herd's three-layer context management exists to prevent.
• Agents run while you work. An agent that pins your only Mac at 100% during a video call is an agent you turn off. Herd's adaptive capacity knows which machines are in a meeting, thermally throttling, or under heavy foreground load, and routes around them.
• Cold loads kill agent latency. A model that is not resident costs 15 to 30 seconds to load. Across a fleet, Herd routes to the machine that already has the model warm.
• One machine cannot hold every model. Agents often want a fast model for routine turns and a heavy one for hard problems. Across a fleet you can keep both resident on different Macs.

OpenClaw vs Claude Code: both, not either

These are different shapes of agent and Herd serves both from the same endpoint.

Running both against one herd is the point: the same pool of Macs serves your terminal sessions and your messaging agent, and the router keeps them from trampling each other. See the Claude Code guide for that side.

Being honest about what Herd does not do

• Herd is not an agent. It does not replace OpenClaw or compete with it. It is the endpoint underneath.
• Herd does not make a small model smart. Routing sends a request to the best machine; it does not change what the model is capable of. If an agent task needs a bigger model, you still need a machine that can hold one.
• Herd does not shard one model across Macs today. It routes many requests across many machines. If you need one model too large for any single Mac, that is exo or Apple's distributed MLX territory.
• Herd has no inbound authentication by default. It assumes a trusted local network. Do not expose it to the internet for an agent to reach remotely; use a private network such as Tailscale.

Related Reading

• Agent Skill, the full ClawHub skill reference
• Claude Code CLI, the other agent Herd is built for
• Adaptive Capacity, how Herd routes around Macs you are actively using
• Quickstart, get the herd running first