Nedster 4.0 Alpha Brings Evidence-Based Local AI Coding to Ollama

Summary

Nedster 4.0 Alpha is a local-first coding and automation agent that sends inference to an Ollama server on Windows, macOS, or Linux, without an account or cloud fallback. Its native tool-calling loop can read, search, edit, and write project files, inspect Git state, run tests, and execute commands, while outcomes are classified from recorded tool evidence rather than the model’s self-report. A change request that changes nothing is marked INCOMPLETE, and an “already done” claim requires current file reads and a passing test run. The project reports that all 16 offline toy-repository checks passed with zero false successes, compared with 1/8 and one false success for the removed v3 engine; five runs were still reported as UNVALIDATED because they lacked trusted post-change validation. When files are changed without tests, auto-verification can run recognized test suites and give the model one repair turn if they fail. Edits require a current SHA-256, are atomic, and can be reversed with /undo when files have not changed since the turn. Command execution is disabled until consent is granted and can use Bubblewrap, Seatbelt, Docker, or opt-in host execution; Windows uses Job Objects for process-tree containment, but native Windows execution does not provide filesystem or network isolation. Nedster supports resumable SQLite-backed sessions, deterministic context compaction, append-only history for Ollama prompt-cache reuse, and hardware-aware model plans that download and configure models during setup. Its documentation reports about 47x faster prefill in one 56K-token measurement with cache reuse, but the evaluation uses small repositories, one local model, and one run per fixture, while the project notes that hostile test code can still spoof simplistic success output. Installation requires Python 3.11+, Ollama, and about 8 GB of RAM; a GPU is recommended but not required. The current status is 4.0 alpha.