OpenCode and Local Models: The Real Power of Free AI Agents

The video opens with a simple question: without paying for any subscription or purchasing any package, how far can a free AI agent take us? To answer this, the open-source, model-agnostic agent harness OpenCode was tested with 11 different models (free, paid-via-API, and locally run) using identical prompts.

OpenCode is a model-agnostic agent harness that runs in the terminal, desktop, or IDE (VS Code, Cursor, Zed, Windsurf). It provides LSP-based error detection, multi-session for parallel agents, shareable session links, and integrations with GitHub, Google, and ChatGPT. Over 75 model providers are supported. The free tier includes the 'Ox Alpha free' model with a quota of 200 queries per 5 hours.

You can add your own API keys to test paid models. OpenAI GPT-5.6, Claude Opus 5, and Gemini 3.7 Flash were used via their APIs. This demonstrates how OpenCode lets you swap models while keeping the same harness, shifting spending from model consumption to harness investment.

For local execution, LM Studio Bionic was used to download Qwen 3.8 B. On a Mac M5 Max with 128 GB unified RAM the model ran entirely in RAM; on Windows a GPU with 24+ GB VRAM is required. The local model was hooked into OpenCode as a 'custom provider'. The model never left the computer and no token fees were incurred.

Tests were performed in four stages: analog clock, event card grid with search/filter, adding sorting to existing cards, and a 3D scroll-driven scene using the 3GS library.

Results: All 11 models produced a working clock (aesthetic differences only). 10/11 rendered the card grid successfully. All 11 added sorting without breaking existing features. 8/11 rendered the 3D scene; three models failed. Free models are functionally on par with paid ones.

When skill files were added, the same model's output changed dramatically. Qwen 3.8 B gained animated transitions, improved colour palette, better typography, and hover effects. Skill files let us steer agent behaviour without changing the model.

LM Studio Bionic serves the model locally via API and can be extended to phones or other computers. Pricing (Sep 2026): MiniMax M3 at $0.30/$1.20 per 1M tokens; GLM-5.3-Flash at $0.075/$0.25; Abacus AI Agent at $20/month. In agentic workflows, token cost is a small fraction; the dominant expense is the harness and data management.

In 2026 a free AI agent is achievable — OpenCode plus a local model yields zero subscription, zero token fees, and data never leaves your device. The real value lies in knowing what you want and expressing it in plain-text files.