Local coding agents become a practical alternative to cloud tools

Local coding agents are becoming more viable as open-weight models improve and harnesses make them useful inside real projects. A local stack combines a model served through an inference engine such as Ollama with a coding harness that can read files, edit code, run commands, and verify changes. The setup is positioned as a transparent, inspectable alternative to proprietary services, with benefits for privacy, offline work, predictable costs, and reproducibility.

Qwen3.6 35B-A3B is the main example, alongside Cohere North Mini Code and Nemotron 3 Nano. Qwen3.6 35B-A3B is about 22 GB to download, requires roughly 30-40 GB of RAM, and can run at usable speeds on systems such as a Mac Mini with M4 or a DGX Spark. For local agent work, anything faster than 20-30 tok/sec is described as reasonable, and both Qwen3.6 and North Mini Code clear that bar in the tests.

The harness matters as much as the model. Qwen-Code is open-source and designed for Qwen models, but Codex and Claude Code can also be connected to Ollama-served models. A small capability benchmark found Qwen3.6 and North Mini Code solving 4 out of 5 tasks in Qwen-Code, while Codex performed even better with Qwen3.6 in the same test set. Claude Code also performed strongly, but used far more tokens on average, while Codex used the least.

Security and privacy remain central concerns because coding agents can read local files and execute commands. Recommended safeguards include auditing the agent codebase, disabling telemetry and usage statistics, avoiding permissive modes such as –yolo, using sandboxes or separate hardware, and treating untrusted repositories as hostile by default.