Reddit user builds Eve, a local autonomous coding agent with a fine-tuned soul

Cloud-based coding agents like Claude Code and Cursor come with tradeoffs: per-token costs, data leaving your machine, and subscription lock-in. To solve this, Reddit user jeffgreen311 created Eve, a local autonomous coding agent that runs on Ollama. Eve has a unique two-layer design. The 'soul' layer uses fine-tuned Qwen 3.5 models (4B at 2.5GB, 8B at 3.4GB) that store personality directly in the weights—not via system prompts. This runs entirely on local GPU and costs nothing per message. The 'worker' layer is cloud-on-demand, using MiniMax M3 (1M token context) for heavy multi-file tasks and Qwen 3.5 397B as a fallback.

Eve uses a three-tier intent router to decide where each message goes: casual to local 4B, light code tasks to local 8B, and complex work to M3. Within a task, Eve executes a 40-round autonomous loop: it writes files, runs bash, reads errors, fixes bugs, and repeats—all while streaming every thought live to a browser. A STEER bar lets users inject mid-task corrections. Eve comes with 16 tools including file operations, git, web search, screenshot, and GUI control. In a real test, Eve passed 9/9 unit tests on the first attempt with a cold prompt. This shows a viable path toward private, self-hosted AI coding assistants with no recurring fees.