wompmacho
887c8a752f
2.2 KiB
title, description, author, date, lastmod, tags
| title | description | author | date | lastmod | tags | ||||
|---|---|---|---|---|---|---|---|---|---|
| Gemini A2A Bridge Setup | Documentation on setting up an Agent-to-Agent bridge to offload Gemini CLI tasks to a local Ollama instance. | wompmacho | 2026-04-12T12:00:00-04:00 | 2026-04-12 |
|
Gemini CLI Local A2A Setup
Setting up an Agent-to-Agent (A2A) bridge allows Gemini CLI to delegate computationally expensive tasks to a local GPU, significantly reducing API token consumption while leveraging high-performance local hardware.
Goal
The primary objective is to offload heavy lifting, such as complex code generation and deep architectural analysis, to a local RTX 4080 GPU located at 10.0.0.109. By routing these requests through a local instance of Ollama, we maintain high inference speed for large models without incurring external API costs.
Bridge Script
A Node.js bridge script acts as the intermediary between Gemini CLI and Ollama. This script is located at /home/wompmacho/a2a-ollama-bridge.js and runs on localhost:8080. It performs the following functions:
- Agent Card Hosting: Serves the required
agent.jsonmetadata at/.well-known/agent.jsonso Gemini CLI can discover the sub-agent's capabilities. - Request Translation: Receives task payloads from Gemini CLI, extracts the prompt, and forwards it to the Ollama API.
- Response Handling: Captures the local model's output and returns it to Gemini CLI in the expected A2A format.
Sub-agent Configuration
The sub-agent is integrated into Gemini CLI via a configuration file located at /home/wompmacho/.gemini/agents/local-gemma.md. This file defines the sub-agent as a remote kind and points to the bridge's metadata endpoint:
---
kind: remote
name: local-gemma
description: "A powerful local sub-agent running Gemma 26B on an RTX 4080."
agent_card_url: http://127.0.0.1:8080/.well-known/agent.json
---
Model Details
The bridge is configured to utilize the gemma4:26b model. This 26-billion parameter model is optimized for the local RTX 4080, providing a balance of reasoning capability and performance that rivals many cloud-based alternatives for specialized technical tasks.