Summary

Let agents delegate short, fast inference tasks to chat-capable models installed on a paired OpenClaw node, with local models auto-discovered from Ollama.

Problem to solve

An agent connected to a capable desktop or server node cannot currently discover and use that machine's local models. Existing Ollama provider support assumes the Gateway can reach Ollama directly, while the existing llm-task tool runs a Gateway-configured model. Neither path provides node-owned, loopback-only local inference.

Proposed solution

Extend the bundled Ollama plugin with two safe node-host commands and one agent tool:

• ollama.models discovers local models from Ollama's loopback API, filters cloud and embedding-only rows, and ranks loaded models first.
• ollama.chat validates an exact discovered model and runs bounded, non-streaming chat on the node.
• node_inference lets agents discover capable paired nodes and run a selected model through the existing authenticated Gateway/node connection.

Keep core plugin-agnostic by using the existing plugin node-host command, node invoke policy, and runtime node APIs. Do not add a new protocol or configuration surface.

Alternatives considered

• Configure Ollama as the Gateway's main model provider. This requires Gateway-to-Ollama network access and changes the agent's primary model rather than delegating a small task.
• Use node exec to call Ollama. This exposes shell execution, lacks typed discovery and filtering, and makes the agent parse provider-specific output.
• Build a new core local-inference protocol. Existing plugin and node seams already carry the required command and result shapes.

Related prior request #2838 concerned first-class local providers on the Gateway; this request is specifically node-owned inference over an already-paired node connection.

Impact

• Affected: users with Ollama models on a desktop or server node.
• Severity: workflow blocker for keeping fast helper inference local to the node.
• Frequency: every attempted delegated local-model task.
• Consequence: today users must expose Ollama to the Gateway network or use shell workarounds.

Evidence/examples

Expected agent workflow:

1. Discover local models on connected nodes.
2. Select a loaded or small model.
3. Run a bounded prompt and receive response, token usage, and timing metadata.

The implementation should prove the real Gateway to paired-node to Ollama path, not only mocked HTTP calls.

Additional information

Ollama's standard node-local endpoint is http://127.0.0.1:11434. Node inference should not reuse remote or cloud Ollama provider configuration.