DEVOURED
Using Local Coding Agents
A practical guide to setting up a fully local coding agent stack that offers transparency and control over your development environment.
What: Sebastian Raschka demonstrates how to use the 'Qwen-Code' harness paired with locally served open-weight models like Qwen3.6 via Ollama. The guide covers performance benchmarking, security auditing for coding agents, and connecting local models to established harnesses like Codex and Claude Code.
Why it matters: Local coding agents provide a privacy-focused alternative to subscription-based proprietary services like Claude or OpenAI, while enabling users to audit the agent's behavior and dependencies.
Takeaway: Clone the Qwen-Code repository, perform a security audit of the source code, and use `ollama pull` to host models locally before connecting your development environment.
Deep dive
- Setup: Use Ollama for efficient local inference; leverage MLX versions if using Apple Silicon.
- Performance: 30-35B parameter models (e.g., Qwen3.6, North Mini Code) provide sufficient performance for coding agent tasks.
- Harnesses: Qwen-Code, Codex, and Claude Code are top options; each requires specific configuration for custom local providers.
- Security: Auditing agent codebases is essential to check for telemetry, network egress, and file access permissions.
- Workflow: Use SSH tunneling if the inference machine (e.g., DGX Spark) is separate from the development machine (Mac).
- Configuration: Use
~/.qwen/settings.jsonto disable telemetry and restrict agent file access.
Decoder
- Inference engine: Software used to run a pre-trained LLM, optimizing it for speed and memory usage on specific hardware.
- Open-weight models: LLMs for which the weights (the learned parameters) are available for public use, unlike 'closed' models whose weights remain hidden behind proprietary APIs.
- Agent harness: The surrounding software environment that gives an LLM the ability to read files, execute shell commands, and verify changes.
Original article
Full article content is not available for inline reading.
Explore all 71 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi