OpenResearch Turns Your Research Directions Into Parallel Agents
alphaXiv's OpenResearch trended on GitHub today and the pitch is narrow enough to be useful: a local-first workspace where each research direction gets its own agent, running in parallel, tracked in a git-native experiment tree. MIT licensed, Rust, and it drives Claude Code, Codex or OpenCode rather than shipping its own model. https://github.com/alphaXiv/OpenResearch
The design decision that matters is the git-native tree. Autoresearch has an obvious failure mode: you fire off eight variations, six of them produce something, and two days later you cannot reconstruct which config produced which number. Putting every experiment in a commit tree means the provenance is structural rather than something you remember to write down. That is unglamorous plumbing and it is the difference between a demo and a tool you still use in a month.
It is genuinely active, not a social spike. The repo shipped orx 0.2.0 on September 11 fixing OpenCode config truncation when output exceeds 64 KiB through a pipe, along with Windows support for the CLI and dashboard and a reworked skill picker, all in the last 48 hours. 326 commits, 1.2k stars, and a desktop app for macOS on top of the CLI. There is also a remote-compute marketplace that aggregates providers, which is either a nice convenience or the actual business model.
The category is filling in fast. hyperresearch trended the same day doing the literature-review version, and the common shape between them is parallel agents plus a local persistent store plus an explicit review pass. Nobody has won this yet, and the useful question when you evaluate one is not how many sources it touches but whether you can tell, three weeks later, where a given claim came from. Get it at https://openresearch.sh
Related reading: https://clauday.com/article/d273430f-5f1d-4622-b7fd-bf5d7bcf4eda
← Back to all articles
The design decision that matters is the git-native tree. Autoresearch has an obvious failure mode: you fire off eight variations, six of them produce something, and two days later you cannot reconstruct which config produced which number. Putting every experiment in a commit tree means the provenance is structural rather than something you remember to write down. That is unglamorous plumbing and it is the difference between a demo and a tool you still use in a month.
It is genuinely active, not a social spike. The repo shipped orx 0.2.0 on September 11 fixing OpenCode config truncation when output exceeds 64 KiB through a pipe, along with Windows support for the CLI and dashboard and a reworked skill picker, all in the last 48 hours. 326 commits, 1.2k stars, and a desktop app for macOS on top of the CLI. There is also a remote-compute marketplace that aggregates providers, which is either a nice convenience or the actual business model.
The category is filling in fast. hyperresearch trended the same day doing the literature-review version, and the common shape between them is parallel agents plus a local persistent store plus an explicit review pass. Nobody has won this yet, and the useful question when you evaluate one is not how many sources it touches but whether you can tell, three weeks later, where a given claim came from. Get it at https://openresearch.sh
Related reading: https://clauday.com/article/d273430f-5f1d-4622-b7fd-bf5d7bcf4eda
Comments