SciForge open-sourced as local-first research agent harness

On October 18, 2026, a developer announced on V2EX the open-source release of SciForge, a local-first Agent Harness for deep learning research, published under the MIT license. According to the post, the project divides the research workflow into four layers. The first is Skill, a set of 24 methodology modules compatible with the DSH and Claude Code specifications. The second is Tool, 46 deterministic capabilities covering literature search, close reading, SSH, experiments, figure generation, and writing. The third is Pipeline, made up of 7 recipes named full, survey, experiment, paper, daily, figure, and litreview. The fourth is GUI, consisting of 8 local web pages that require no build step and no CDN.

The stack is built on Python 3.12, Typer, FastAPI, and a native frontend. By default, SciForge connects to deepseek-flash and deepseek-v4-pro, and it also supports OpenAI-compatible APIs. The author explained that the motivation was that while LLMs have become strong at writing code, research workflows remain fragmented, with literature in one tool, experiments over SSH, and plotting in another script. The project ships with 355 offline tests, and its CI covers Python 3.10 and 3.12. The post includes a GitHub repository link and invites stars, issues, and pull requests. Public information currently stops at this open-source launch, and the available material contains no reports of later developments.