Aider + qwen2.5-coder:14b as a local Claude Code alternative
What they built
The author uses Claude Code and wanted to know what the self-hosted version feels like: no API, no subscription, code that never leaves the machine, and no dependence on a provider's prices or survival.
He paired Aider (a CLI that reads and edits project files) with a local coder model and recorded building a terminal tic-tac-toe game.
Setup
- Ubuntu 24.04, Ryzen 9 9900X, 64 GB RAM, RTX 5060 Ti 16 GB
- Ollama in a container (nerdctl compose) with CUDA and flash attention
- qwen2.5-coder:14b, preloaded with a 3-hour keep-alive
- Aider v0.86.2 in a Python 3.13.9 virtualenv
Results & numbers
- The core game logic was correct on the first pass
- Input validation and unit tests needed several rounds of prompting
Gotchas
- Out-of-range input was reported as 'already taken' even though the prompt asked for validation
- Generated unit tests had a bug
- Sometimes it's faster to edit the file yourself than keep prompting
“Sometimes the most efficient path with AI coding tools is knowing when to stop prompting and start typing”— mrmcmuffinz.dev
Do it yourself
- Run Ollama with GPU access and pull qwen2.5-coder:14b
- Preload the model with a long keep_alive so it stays warm
- Install Aider and point it at the local Ollama endpoint
- Write a detailed prompt (features, validation, tests) and start in ask mode
- Review and test everything it writes
curl http://localhost:11434/api/generate -d '{"model":"qwen2.5-coder:14b","keep_alive":"3h"}'Related
Home RAG assistant for embedded-systems docs on Proxmox
Ollama, llama.cpp, Open WebUI, BGE-M3 and SearXNG in a GPU-passthrough VM. Moving from 6 GB to 16 GB of VRAM changed which models could run.
Hybrid workflow: local for private and repetitive, cloud for hard
A Linux sysadmin stopped trying to replace cloud LLMs. Ollama now handles private docs, overnight article summaries and VS Code autocomplete.
Offline RAG over 268 Obsidian notes, queryable from a coding agent
nomic-embed-text, ChromaDB and mistral-nemo:12b give cited answers from a vault of internal docs, and an MCP server lets the coding agent search it too.
14 months local-first: what still breaks
An honest report from a three-machine home lab. Local handles about 85% of tasks, but four kinds of work still fail, and that cut cloud spend from $187 to about $25 a month.