Personal AI Agent
Setup Guide

OpenCode + Cloudflare Workers AI + AI Gateway + Google Workspace MCP

Free plan • sean.connellan@gmail.com

Architecture

How It All Connects

Your Laptop

OpenCode — Terminal AI client

Connects to MCP servers for tools, sends LLM requests through AI Gateway

Local MCP

Google Workspace — Gmail, Calendar, Drive, Docs, Sheets, Slides, Chat

Runs locally, OAuth credentials stay on your machine

Cloudflare

AI Gateway — Caching, rate limiting, logging, DLP

Sits between OpenCode and Workers AI, saves neurons

Workers AI

Free LLM Models — Llama, Qwen, Granite, Gemma

10,000 neurons/day on free plan, OpenAI-compatible API

Data Flow

You type a prompt in OpenCode

OpenCode sends the request through AI Gateway (cached? logged? rate-limited?)

AI Gateway forwards to Workers AI for inference

If the AI needs a tool, OpenCode calls the Google Workspace MCP server

MCP server uses OAuth to call Google APIs, returns results to the AI

What is OpenCode?

A terminal-based AI assistant that orchestrates everything:

Chat interface for LLM interactions
Connects to any LLM provider (not locked to one)
Discovers and calls MCP server tools
Loads skills for specialised workflows
Handles file editing, bash, web fetch
Permission system (allow/ask/deny)

Install

# macOS (Homebrew)
brew install anomalyco/tap/opencode

# Or via curl
curl -fsSL https://opencode.ai/install | bash

Verify

opencode --version
opencode          # launch TUI

What is MCP?

Model Context Protocol — a standardised way for AI models to discover and invoke external tools. Think of it as "USB for AI".

Local MCP

stdio transport

Runs on your laptop as a child process. Google Workspace MCP uses this — your OAuth credentials never leave your machine.

Remote MCP

Streamable HTTP transport

Runs on Cloudflare Workers. Accessible from anywhere, can be shared, scales to zero. Optional for personal use.

AI Gateway — Your AI Control Plane

Caching

Identical prompts served from cache — zero neurons consumed

Rate Limiting

Prevent runaway usage from burning your free allocation

Analytics

Track every request — tokens, latency, cost, errors

DLP

Redact sensitive data before it reaches the LLM

Logging

100K logs stored free — audit what your AI is doing

Model Routing

Switch between models/providers without changing client config

Workers AI — Free LLM Models

10,000 neurons/day on the free plan. Serverless GPU inference on Cloudflare's global network.

Model	Size	Neurons/M input	Best for
`llama-3.1-8b-instruct-fp8-fast`	8B	4,095	General purpose (recommended)
`qwen3-30b-a3b-fp8`	30B	4,577	Reasoning tasks
`llama-3.3-70b-instruct-fp8-fast`	70B	26,668	Complex tasks
`granite-4.0-h-micro`	Small	1,542	Quick tasks
`gemma-4-26b-a4b-it`	26B	~4,500	General purpose

8B model: ~2.4M input tokens/day free • 70B model: ~375K input tokens/day free

Setup Steps

Step 1 — Cloudflare Account

Create Account

Sign up at dash.cloudflare.com/sign-up
Use sean.connellan@gmail.com
Free plan — no credit card

Find Account ID

Dashboard → Workers & Pages → Account ID in sidebar

Create API Token

Go to dash.cloudflare.com/profile/api-tokens
Use "Edit Cloudflare Workers" template
Add: Workers Scripts Edit, Workers AI Edit, Account Settings Read
Copy token — shown only once

~/.zshrc

export CLOUDFLARE_ACCOUNT_ID="your-id"
export CLOUDFLARE_API_KEY="your-token"

Step 2 — AI Gateway

Create Gateway

Dashboard → AI → AI Gateway
Click Create Gateway
Name: personal-ai-gateway

Your gateway URL:

https://gateway.ai.cloudflare.com/v1/
  {ACCOUNT_ID}/personal-ai-gateway/
  workers-ai/v1

Recommended Settings

Setting	Value
Caching	Enabled
Rate Limiting	50 req/min
Logging	Enabled
DLP	Optional

Free: 10 gateways, 100K logs, core features included

Step 3 — Google Cloud OAuth

Google Cloud Console

Create project at console.cloud.google.com
Enable APIs: Gmail, Calendar, Drive, Docs, Sheets, Slides, Chat, People, Tasks
OAuth consent screen → External (required for @gmail.com)
Add sean.connellan@gmail.com as test user

Create Credentials

APIs & Services → Credentials → OAuth client ID
Type: Desktop Application
Download JSON → save to ~/.google-mcp/credentials.json

"This app isn't verified" warning is expected in Testing mode. Click Advanced → Go to app.

Step 4 — Google Workspace MCP Server

Install

git clone https://github.com/gemini-cli-extensions/workspace.git \
  ~/projects/google-workspace-mcp
cd ~/projects/google-workspace-mcp
npm install && npm run build

First Auth

Run node workspace-server/dist/index.js --debug — browser OAuth opens automatically

Auth Management

node scripts/auth-utils.js status / clear / expire

Provides 50+ tools: Gmail, Calendar, Drive, Docs, Sheets, Slides, Chat — all via local stdio transport

OpenCode Config

OpenCode — LLM Provider Config

~/.config/opencode/opencode.jsonc

{
  "model": "cf-gateway/@cf/meta/llama-3.3-70b-instruct-fp8-fast",
  "small_model": "cf-gateway/@cf/meta/llama-3.1-8b-instruct-fp8-fast",
  "provider": {
    "cf-gateway": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Cloudflare AI Gateway",
      "options": {
        "baseURL": "https://gateway.ai.cloudflare.com/v1/{ACCOUNT_ID}/{GATEWAY_ID}/workers-ai/v1",
        "apiKey": "{env:CLOUDFLARE_API_KEY}"
      },
      "models": {
        "@cf/meta/llama-3.3-70b-instruct-fp8-fast": { "name": "Llama 3.3 70B" },
        "@cf/meta/llama-3.1-8b-instruct-fp8-fast": { "name": "Llama 3.1 8B" }
      }
    }
  }
}

OpenCode — MCP Servers & Permissions

~/.config/opencode/opencode.jsonc (continued)

"mcp": {
  "google-workspace-local": {
    "type": "local",
    "command": ["sh", "-c",
      "cd /Users/sean/projects/google-workspace-mcp && node scripts/start.js"]
  },
  "sean-mcp-server": {
    "type": "remote",
    "url": "https://sean-mcp-server.<subdomain>.workers.dev/mcp"
  }
},
"permission": {
  "edit": "allow",
  "webfetch": "allow",
  "bash": {
    "*": "allow",
    "rm *": "ask",
    "rm -rf ~*": "deny",
    "sudo *": "ask"
  }
}

Skills — Teach the AI Specialist Workflows

SKILL.md files in ~/.config/opencode/skills/<name>/

Loaded dynamically when a task matches their description. Turn a general-purpose AI into a specialist.

Example Skills

email-summariser — Prioritised email digest
daily-briefing — Calendar + email + tasks
call-notes-summary — Structured call notes

email-summariser/SKILL.md

---
name: email-summariser
description: Summarises unread emails
---

## Workflow
1. gmail_search: is:unread newer_than:1d
2. gmail_get: fetch each email
3. Categorise: URGENT / ACTION / FYI
4. Present structured summary

10,000 Neurons/Day
Is Plenty

~2.4M input tokens with Llama 3.1 8B • ~375K with Llama 3.3 70B

Hit the limit? Workers Paid is $5/month for unlimited neurons at $0.011/1K

Full guide: github.com/gemini-cli-extensions/workspace • opencode.ai • developers.cloudflare.com/agents

Slides PDF Full Tutorial PDF

Personal AI AgentSetup Guide