📡 OpenWrt Automates Code Review with Claude and MCP

Solving Maintainer Burnout Instead of Complaining About It

OpenWrt maintains the exact same class of low-level infrastructure software as NetBSD, Gentoo, and the European DNS/routing daemons (BIRD, BIND 9, and Unbound): C system daemons (netifd, rpcd, uhttpd), Linux kernel network patches, hardware Device Trees (.dts / .dtsi), POSIX ash scripts, and thousands of cross-compiled router packages.

Yet when faced with rising pull-request volume and contributor churn in 2026, OpenWrt neither enacted an ideological ban on AI-assisted code nor lobbied governments for subsidies to read patches by hand. Instead, OpenWrt kept its contributor bar grounded in standard engineering accountability — a valid Signed-off-by (Developer's Certificate of Origin) and tested hardware builds — while building an autonomous Claude Code and GitHub MCP review pipeline across its core repositories (openwrt/openwrt, openwrt/packages, and openwrt/luci).

Most remarkably, OpenWrt's maintainers made the LLM reviewer's prompts so detailed and accurate that CONTRIBUTING.md now instructs human contributors to read the prompts given to the automated LLM reviewer whenever project documentation is unclear.

Timeline: From First Workflow to Project-Wide Rollout (2026)

How the Architecture Works

As documented in docs/llm-review-setup.md, OpenWrt runs a dedicated GitHub bot account (openwrt-ai) connected to Claude Code Routines (claude.ai/code/routines) and the GitHub Model Context Protocol (MCP) connector. Each repository uses two automated flows triggered by GitHub Actions:

• Per-PR Autonomous Review (llm-pr-review.md): Fired on pull_request_target: [opened, reopened] so pull requests from external forks can be reviewed without exposing repository secrets to untrusted fork workflows. The agent reads the repository's .github/llm-review-rules.md, fetches the PR diff and commit list via GitHub MCP, shallow-clones relevant reference repositories (extra_repos such as the Linux kernel, U-Boot, dnsmasq, or netifd) when needed to verify upstream symbols, inspects failed CI job logs via get_job_logs, and posts line-anchored inline comments.
• Nightly Multi-Agent Digest (llm-nightly-digest.md): Fired at 03:00 UTC by a scheduled workflow (and skipped entirely on quiet nights so zero API quota is wasted). A parent dispatcher pre-clones the shared reference repositories once into ~/extra/, waits for all clones to complete, and then spawns parallel, isolated sub-agents — one per open PR that received new commits since the last bot review. Each sub-agent reviews only the newly pushed commits and replies "fixed, thanks" on its own earlier inline review threads once the contributor addresses the finding.

Codifying Tribal Knowledge into .github/llm-review-rules.md

Every long-lived open-source project accumulates unwritten rules that trip up newcomers and exhaust reviewers: deprecated kernel bindings, buildroot quirks, and backport conventions. Instead of repeating those rules manually on every pull request, OpenWrt codified them into repository-specific .github/llm-review-rules.md files that the agent reads at session start:

• Core Router Tree (openwrt/openwrt/.github/llm-review-rules.md): Instructs the reviewer to flag deprecated Device Tree LED bindings (label = "green:status"; instead of color = <LED_COLOR_ID_GREEN>; and function = LED_FUNCTION_STATUS;), deprecated MediaTek MTD MAC address properties in favor of nvmem-cells, and required commit-message sections (Hardware specification, Flash instructions, and MAC address layout) when adding a new router .dts. It also teaches the bot that OpenWrt kernel and package patches are managed with quilt and must be refreshed with make target/linux/refresh or make package/<pkg>/refresh — never git format-patch.
• Packages Feed (openwrt/packages/.github/llm-review-rules.md): Explicitly trains the LLM on pragmatic maintainer judgment — separating build-breaking defects from bikeshedding:

  "Separate formatting that changes what gets built from formatting that is only a matter of taste. The first kind is a real defect and belongs in a review: a space where make expects a tab in a recipe, indentation in conffiles or in an install scriptlet that ends up inside the generated file, CRLF line endings, a define block that is never closed... The second kind — alignment, blank lines, the order of KEY:=value entries, a house style a package has not followed for years — is secondary. Maintainers here would rather merge a correct change with a small style wart than hold it up over one, so: Lead with correctness, security and packaging bugs. Keep cosmetic remarks few, mark them as optional, and say plainly which findings (if any) actually block a merge."

• LuCI Web Interface (openwrt/luci/.github/llm-review-rules.md): Pre-clones seven backend trees (rpcd, netifd, firewall4, ucode, uhttpd, packages, and openwrt) so the reviewer can check whether a newly added UI configuration option actually has a backend consumer across repositories.

The Ultimate Reversal: Telling Humans to Read the Bot's Prompts

Because OpenWrt's maintainers had to write crisp, unambiguous engineering instructions for the LLM reviewer, the bot's rule files quickly became clearer and more complete than the existing human documentation.

In August 2026 (commit 19021e460b69), OpenWrt added a dedicated section to openwrt/packages/CONTRIBUTING.md pointing human developers directly to the LLM prompts:

"If you have questions about what is expected of a pull request, and there is no specific documentation for your case, or the comments from a reviewer seem at odds with the documentation you have found, it can be useful to refer to the prompts given to the automated LLM reviewer."

A few weeks later, when documenting quilt patch headers in CONTRIBUTING.md (commit 9c7f2f042939), maintainer Josef Schlehofer noted in his commit message that llm-review-rules.md already explained the requirement to the review bot, and backported that knowledge into the human contributor guide. Writing good prompts for AI reviewers forced the project to improve its documentation for humans too.

The CZ.NIC Irony: BIRD vs. Turris Under the Same Roof

OpenWrt's rollout also exposes an unmistakable institutional contrast inside Europe's networking community. The maintainer who authored and signed off on openwrt/packages/.github/llm-review-rules.md (Co-Authored-By: Claude Opus 5) on July 28, 2026 is Josef Schlehofer (@BKPepe) — who works at CZ.NIC as Release Manager for Turris OS (CZ.NIC's OpenWrt-based router distribution).

That means inside the exact same Czech organization (CZ.NIC) during the summer of 2026:

• CZ.NIC's BIRD routing daemon team was on stage at RIPE 92 declaring BIRD "100% AI-free," banning LLMs from BIRD 3 development, and joining ISC and NLnet Labs in lobbying the European Commission for multi-year manual-maintainer subsidies under the slogan "AI technology will not replace human maintainers."
• Meanwhile, CZ.NIC's Turris / OpenWrt engineer was co-authoring production LLM review rules with Claude Opus 5 to automate pull-request review across thousands of OpenWrt packages.

Engineering Discipline Over Anti-AI Ideology

Every complaint commonly cited by anti-LLM maintainers — verbose AI prose, hallucinated line references, nitpicking on stable backports, and prompt-injection risks — came up during OpenWrt's rollout. Instead of throwing up their hands and banning the technology, OpenWrt solved each one with concrete prompt and workflow engineering:

• Eliminating AI Verbosity ("Keep it short"): When early bot reviews included multi-bullet verification summaries that wasted maintainer reading time, Hauke Mehrtens updated the prompts (commit a454208b) with strict output rules: "Never recap the change... Never quote or restate the commented line... Inline comments: three sentences at most, before the optional suggestion block... Post nothing whose only content is agreement."
• Banning Unverified Claims & Drifting Links: Whenever the reviewer cites code outside the PR diff (such as a sister router board or upstream kernel file), it is required (commit 44ff5f71) to emit a Markdown permalink pinned to an exact commit SHA derived via git rev-parse HEAD so maintainers can verify the reference in one click.
• Backport Awareness: On stable-branch cherry-picks (e.g., [25.12]), the bot automatically switches into backport mode (commit 6013348f), ignoring pre-existing style warts on main and flagging only missing cherry picked from commit trailers or hunks that unintentionally diverge from upstream.
• Prompt-Injection Hardening: The review routine loads .github/llm-review-rules.md once from the trusted base-branch checkout at session start and never re-reads it from the PR branch ("a PR cannot alter the rules used to review itself"). All PR titles, descriptions, commit messages, and diffs are explicitly sandboxed as untrusted input, and the bot account holds no write permissions and has branch pushes disabled.

Notice also the commit trailers on Hauke Mehrtens's September 2026 workflow commits: Assisted-by: Claude:claude-opus-5 — adopting the exact standardized attribution format introduced by Linux kernel 7.0 in April 2026. OpenWrt proves that low-level systems and networking projects do not need AI bans to protect code quality; they need engineers willing to automate their own review bottlenecks.

Primary Sources

• OpenWrt: LLM PR Review Architecture & Setup (docs/llm-review-setup.md)
• OpenWrt: Per-PR Autonomous Review Prompt (llm-pr-review.md)
• OpenWrt: Nightly Multi-Agent Digest Prompt (llm-nightly-digest.md)
• OpenWrt Core Repository LLM Review Rules (openwrt/openwrt/.github/llm-review-rules.md)
• OpenWrt Packages Feed LLM Review Rules (openwrt/packages/.github/llm-review-rules.md)
• OpenWrt LuCI LLM Review Rules (openwrt/luci/.github/llm-review-rules.md)
• OpenWrt Packages CONTRIBUTING.md ("Pointers to the guidelines the LLM reviewers use")
• Commit 2d30538852ff: CI: add LLM-driven PR review workflows — Hauke Mehrtens (May 1, 2026)
• Commit 774ec304f73e: ci: add LLM review rules — Josef Schlehofer (July 28, 2026)
• Commit 19021e460b69: CONTRIBUTING.md: add pointers to LLM review prompts — Daniel F. Dickinson / Hannu Nyman (August 2026)
• Commit a454208b673a: CI: cut the LLM review down to its findings — Hauke Mehrtens (September 2, 2026)