5 June 2026
Copilot bills, Gemma 4 & MCP tutorial
Build your own AI coding workbench for the price of electricity — then find out why Copilot users are rage-quitting their subscriptions.
Hello,
Two posts this week. The first looks at the macro picture: the week’s major AI signals, including what happens when the subscription model finally buckles under its own inference costs. The second is a hands-on guide to escaping that exact subscription fatigue by building your own local AI coding setup from scratch.
GitHub Copilot AI Credits & Claude Opus 4.8 https://daehnhardt.com/blog/2026/06/05/github-copilot-ai-credits-claude-opus-4-8/
June arrived, and with it, GitHub Copilot’s meter started running. One AI Credit equals $0.01, and Pro+ users have reported burning through noticeable chunks of their monthly allowance in ordinary coding sessions. Think of it like a taxi that now charges by the metre rather than a flat fare to the airport — useful information to have before you let an agent helpfully rewrite the wrong component three times whilst trying to fix a pagination bug.
On the exact same day, Anthropic announced that it had confidentially filed a draft S-1 with the SEC, formally kicking off what is shaping up to be quite the IPO race against OpenAI. Market reports put its annualised revenue run rate at $47 billion. Public market investors are, generally speaking, considerably less forgiving of “we’ll sort monetisation eventually” than private ones — so we may soon learn more about the actual economics of frontier AI than any press release has ever voluntarily disclosed.
Elsewhere this week: Claude Opus 4.8 ships dynamic workflows for parallel agentic systems — spawning hundreds of subagents with individual context windows, then aggregating results coherently, which is the architectural shift that moves agentic work from impressive demo to something you would actually run in production. MiniMax M3 becomes the first open-weight model combining frontier-level coding with a one-million-token context window.
Gemma 4 12B brings multimodal agents to 16GB of laptop RAM under Apache 2.0. Microsoft quietly declares partial independence from OpenAI with seven homegrown MAI models at Build 2026. And OpenAI moves GPT-Rosalind toward biodefence workflows — offering it to vetted developers and government partners for pandemic preparedness and biosurveillance work — which, cheekily, introduces indirect prompt injection as a first-order security concern in contexts where the stakes are rather higher than a misconfigured dashboard.
Local AI Agents with Cline, Ollama, and MCP https://daehnhardt.com/blog/2026/06/04/local-ai-agents-cline-ollama-mcp/
If GitHub’s new usage-based billing has you looking for alternatives, now is the perfect time to move your agentic workflows locally. I just published a new step-by-step tutorial on how to wire up Cline with Ollama and the Model Context Protocol (MCP). Pair this setup with the newly released Gemma 4 12B, and you have a frontier-grade coding assistant running entirely on your laptop — no API keys or metered bills required.
Think of MCP — the Model Context Protocol — like a USB port for AI: instead of a model that can only talk, you get one that can actually do things. Without tools, asking an AI about your codebase is rather like asking a well-read stranger to fix your plumbing from memory. With MCP connected, it can read your actual project files, query your database, and search your documentation — not a hypothetical project it has seen in training, but yours.
The setup I have written about combines Cline (an open-source VS Code extension), Ollama (which runs open models directly on your own machine), and an MCP server that wires them together. Cline is free; Ollama is free; the only bill is your laptop’s heat and RAM, which at least has the virtue of being visible, unlike the background token burn of certain cloud tools.
The companion project includes a working semantic file search tool — a 274 MB embedding model via Ollama, entirely offline — with both stdio and FastAPI transports, unit tests, and a walkthrough of how MCP discovery directories actually work. There is also a frank section on security, because “runs locally” does not mean “runs safely,” and a note on what it genuinely takes to publish an MCP server responsibly.
Warm wishes, Elena
Enjoyed this? Get more like it.
Weekly notes on AI tools, Python, and what I'm actually building — plus two free gifts: the 15-page Fantastic AI: The 2026 Toolkit and a Git Commands & Contribution Workflow Cheatsheet.