NVIDIA's Boro brings Sashiko-style AI kernel review to your local tree

NVIDIA has an open-source tool for AI-assisted Linux kernel work called Boro. It is a command-line program written in Rust, licensed Apache-2.0, and the repository dates from June 2026. Andrea Righi of NVIDIA presented it at the Linux Plumbers Conference in early October under the title “Boro: Do We Need More AI-assisted Kernel Tooling?” Boro is not new this week, but the talk put it in front of kernel developers, so this is a catch-up.

Where it fits next to Sashiko

The README says Boro “is inspired by Sashiko” and that “the two tools share the same underlying review prompts, but target different moments in the patch lifecycle.” Sashiko, which we covered through its own LPC numbers, reviews patch series after they appear on public mailing lists. Boro is for “interactive local review and repair while the developer is still shaping the series or carrying it downstream”: backports, distro kernel maintenance, security-fix integration and local patch stacks that may never reach the list.

The LPC abstract says the talk will “open a discussion on whether some of its functionality could be integrated into Sashiko.”

What it does

• boro review RANGE runs a multi-stage agentic review with kernel-focused prompts and ends with an LKML-style summary.
• boro build RANGE checks out each commit in its own worktree, builds it with vng -b, and has the model triage the build log.
• boro test RANGE also boots each kernel under virtme-ng and runs a model-picked quick test, such as a matching kselftest. A --plan flag writes a test plan without running it.
• boro apply cherry-picks a commit range, or a series fetched from lore with b4, and proposes conflict resolutions that a second model must approve, up to 10 rounds per hunk.

The README argues that feeding “real compiler output and real kernel runtime output back to the model, not just the diff text” is what justifies running locally.

Two models, one budget

Boro splits the work between a cheap model (BORO_MODEL) for broad and specialist discovery and a stronger one (BORO_VALIDATION_MODEL) for the baseline review, validation and the report. Token use is broken out per stage. It talks to any OpenAI-compatible endpoint, so a local model can do the bulk of the reading, and it can also drive the Claude, OpenCode or Codex CLIs through --backend.

The repository is small, with a few dozen stars, and its last push was in September. Install it with cargo install --path . from a checkout; build and test need virtme-ng.