about bernstein.run
who
bernstein.run is the project site for bernstein, an open-source orchestrator for cli coding agents (claude code, codex, cursor, aider, gemini cli, and ~40 more). it is built and maintained by alex chernysh as a solo project, self-bootstrapped, no investor money. the site, the python package on pypi, the npm wrapper, and the docs all ship under apache 2.0. public profiles: github.com/chernistry and x.com/@alex_chernysh.
why this project exists
the short version is in the project readme: i was paying about $400/month in claude bills running three coding agents in parallel and getting nondeterministic merges back. the bill was annoying, the merges were worse. bernstein is the orchestrator i wished i had then - a deterministic python scheduler in front of whichever cli coding agents you already trust, with parallel git worktrees, a quality-gate stage, and an hmac-chained audit log so the run is repeatable and the artefacts are accountable. it is not a magic prompt; it is a scheduler.
what qualifies the voice
concrete signals rather than years-of-experience filler. bernstein ships ~40 cli adapters (claude code, codex, cursor, aider, gemini, openai agents sdk, and the rest - see /vs for the comparison index), an hmac-signed audit log, sigstore-style lineage records per artefact, a contextual-bandit router for cheapest-passing model selection, and an mcp server so the orchestrator itself is callable from any mcp client. architectural decisions are public: parallel worktrees over shared state, file-based session store over database, sub-process isolation over in-process plugin loading. the source is on github and the api docs are on readthedocs; the design rationale lives in commit history, prs, and /blog rather than in marketing copy.
how content is sourced
every blog post carries a publish date in its frontmatter and is dated in the visible header. comparisons (the /vs hub, the /why-bernstein decision page, and the per-adapter pages under /vs/<adapter>) cite the upstream project's own documentation as the source of feature claims wherever the claim is about the other tool; where we test a claim ourselves, the test command is published in the post body so it can be re-run. benchmark numbers come from a reproducible eval harness (the benchmark page includes the methodology and a re-run script). model prices on /cost are dated and link to the upstream price page; the date the table was captured is printed on the page.
bias disclosure. i build bernstein, so any claim on this site about bernstein vs another tool is biased by that authorship. i try to keep the bias visible rather than hidden: the benchmark page publishes a 10-task suite where bernstein loses 4 of 10 against the comparison set, the /why-bernstein page includes explicit “who this is not for” sections, and the comparison source memo is the /vs hub itself - each comparison page links the upstream tool's docs at the top so a reader can audit the claim against the source. if a comparison reads as marketing rather than as a fair assessment, file an issue on the bernstein repo and i'll either fix the copy or take the page down.
contact
email forte@bernstein.run, github issues on the bernstein repo, or x.com/@alex_chernysh for short notes.