Local Code Assistant
Copilot-class convenience, on your machine.
Continue + Ollama turns VS Code into a coding assistant that runs entirely locally: chat, edit, and tab autocomplete, with your code never leaving the drive. No Copilot/Cursor subscription, no per-seat cost, no code uploaded to a vendor.
The honest part first
Local coding models (7-32B) are currently weaker than the best cloud models on hard tasks. This kit does NOT claim to be "as good as Copilot". Its pitch is privacy and cost: the everyday 80% of coding assistance works great, for $0/month, and client/NDA code stays on your machine. For the hard 20% you keep one cloud seat (see guide page 16).
What's included
- 📄 The Guide (20-page PDF) — architecture (editor <-> Continue <-> Ollama), 6-step setup, model selection by VRAM, the 5 prompts explained, tuning, troubleshooting, honest limits vs cloud
- ⚙️ config.yaml — current Continue v1 schema, wired to Ollama: chat + edit + a separate autocomplete model
- ⚙️ prompts.md — 5 real reusable coding prompts (explain, tests, refactor, fix, commit)
- 📊 XLSX (2 tabs) — model index by VRAM tier + config reference
- 🖨️ Cheatsheet (2-page PDF) — setup + model pairing + slash commands
Model recommendation by VRAM
- 8GB — chat: qwen2.5-coder:7b (4.7GB) + autocomplete: qwen2.5-coder:1.5b
- 16GB — chat: qwen2.5-coder:14b (9GB) + autocomplete: qwen2.5-coder:3b
- 24GB — chat: qwen2.5-coder:32b (20GB) + autocomplete: qwen2.5-coder:3b
- No GPU — qwen2.5-coder:1.5b, works but slow
All model tags verified on ollama.com/library.
Why trust this?
The config is the current Continue v1 format, verified against docs.continue.dev; the model tags and sizes are verified against the Ollama library. No invented benchmarks — the guide says exactly where local models are weaker than cloud.
Who this is for
Developers and freelancers who want editor-integrated AI on local models: no Copilot/Cursor subscription, no code leaving their machine, works offline.
Who it's NOT for
People who want frontier-model parity on the hardest tasks, or who don't want to run Ollama/VS Code. Cloud Copilot/Cursor is genuinely better for the hard 20% — keep it for that.