Basecamp says a team of Claude Code agents rewrote its Campfire chat app in Rust in about a day, and the port is now 95 times faster on the room-page test

Basecamp has published a Rust port of Campfire, its self-hosted chat product, and says on the repository page that the port was "built in about a day by coordinated Claude Code agents, each owning one crate or harness component". The repository, basecamp/once-campfire-rust, went up on 26 September 2026 under the MIT licence and was still being pushed to on the morning of 30 September, with v0.1.2 as the latest tagged release.

The port replaces the Ruby, Puma, Redis, Resque and Thruster stack of the original with a single Rust binary that reads the same SQLite database and storage directory, so a live Campfire install can move over without moving its data. It keeps the same UI. That "same UI" claim is what the repository's parity gate is for: the README says 5,018 UI cells are compared pixel for pixel against the Rails app on a fixed build, and every cell has to pass before a change is accepted.

The benchmarks the repository actually publishes

The performance numbers on the repository page are the Basecamp team's own, run on 4 pinned CPU cores with host networking against pinned containers, using three interleaved runs per app on identical seed data and a frozen clock. On those runs the Rust room page serves 20,213 requests a second and the Rails one serves 212, which is the 95 times figure. Real-time messaging to 10,000 connected clients reaches 656,183 deliveries a second on the Rust build and 9,485 on the Rails one, a 69 times difference. A cold start takes 143 milliseconds on the Rust binary and 2,607 on Rails. Idle container memory sits at 15 megabytes against 355.

Two things about the method deserve a plain statement. The tests run inside pinned containers on the same host, so the numbers compare the two software stacks under the same conditions rather than describing any specific production environment. And the benchmarks are the Basecamp team's own; a third party has not repeated them at the time of writing.

What the "coordinated agents" claim actually points at

The interesting part for a team thinking about AI code review is what sits alongside the code. The README points to plans/rust-conversion.md, the plan the agents followed, and plans/overnight-report.md, the write-up of one unattended overnight run. It says the plan was reviewed by "Codex" before the agents ran it, and that each optimisation after the first build is a separate pull request with its measurements committed under bench/results/.

That is more evidence than most "we built X with AI" posts leave behind: a plan the agents worked from, a report of what they did while nobody was watching, a parity gate that every change has to clear, and a benchmark run committed with each pull request. A team trying to write an eval suite for AI-written code has, in this one repository, a working example of one particular answer to that problem: parity tests as the gate, benchmarks committed with the change, and the plan the agent worked from left in the repository so it can be read after the fact.

Source

Repository: basecamp/once-campfire-rust on GitHub. All performance numbers and the "one day" and "Claude Code agents" claims come from the project's own README, and are not independently verified here.