local-llm-agentic-coding

An installable agent skill that sets up a free, local, offline LLM (via Ollama) and connects Claude Code or OpenCode to it — so your coding agent runs on a model you own, with no API bill and no data leaving your machine.

It also does the thing most "run Claude Code for free" guides skip: an honest hardware reality-check. Making a local agent work and making it usable are two different walls, and this skill tells you up front which side of them your box is on.

Companion to the crash course "Local AI: Agentic Coding on a Model You Run" in The AI Agent Factory. The course walks a beginner through installing this skill and letting their agent drive it; this repo is the skill that does the work.

Install

Into Claude Code and OpenCode at once:

npx skills add panaversity/local-llm-agentic-coding --agent claude-code opencode -y

List the skill's exact name first if you want to be sure (the installer silently drops names it doesn't recognize):

npx skills add panaversity/local-llm-agentic-coding --list

Then just tell your agent what you want — "set me up to run Claude Code on a local model" — and it will follow the skill: reality-check the box, install Ollama, tune a model, wire the tool, and verify end to end.

Note on discovery: installing by panaversity/local-llm-agentic-coding works the moment this repo is public. Appearing in the skills.sh directory is separate — that listing is driven by real install activity, not by publishing, so it shows up there once people actually install it.

What's inside

File	What it does
SKILL.md	The instructions your agent follows.
scripts/check-hardware.sh	Honest verdict: usable agent, or learning rig?
scripts/install-ollama.sh	Lean, user-space Ollama install (~112 MB on a CPU VPS).
scripts/make-model.sh	Pull + tune a model with num_ctx large enough for a coding agent's prompt.
scripts/launch.sh	Point Claude Code at the local model (Anthropic endpoint + env vars).
scripts/launch-opencode.sh	Point OpenCode at the local model (OpenAI endpoint + opencode.json).
scripts/verify.sh	Prove chat + a real tool call work, and benchmark per-turn speed.
references/	Hardware sizing, the two-door wiring, troubleshooting, other backends.

The honest part

Two walls, and cheap hardware usually misses both:

• Capability — the model must emit valid tool calls turn after turn. Fixed by a bigger model (~14B floor, 30B-class MoE sweet spot), not faster hardware.
• Throughput — the box must chew the agent's long system prompt fast enough. Fixed by a GPU, not a smarter model.

On a laptop with no GPU this is a superb learning rig — the wiring, the API, the agent loop all provably work — but not a daily driver. The setup is identical on a GPU you own or rent; only the model tag and the speed change. See references/hardware-sizing.md.

License

MIT — see LICENSE.