All capabilities Building software

Work through an entire repository, not a pasted file

An agent that reads the codebase itself, edits across many files, runs the commands, reads the failures, and opens the pull request.

DocsOfficially documentedTestNot tested by benchr Medium · Agent / CLI Docs checked September 1, 2026

Claude CodeClaude Opus 5Claude Sonnet 5Claude Opus 4.8

Overview

Reads and edits files across a project, runs commands, and works with git directly - staging changes, writing commit messages, creating branches, opening pull requests.
Runs on the terminal, in VS Code and JetBrains, in a desktop app, and in the browser, against the same engine and the same project settings.
Takes piped input, so logs or a changed-file list can be fed straight in from the shell.

01Put the project's rules in a CLAUDE.md at the repository root - it is read at the start of every session.
02Ask for the whole loop, not the diff: write the change, run the tests, and fix what fails.
03Package anything you repeat as a skill, so the same procedure is one command next time.
04Use hooks for the non-negotiables - format after every edit, lint before every commit - rather than restating them in prose.

Most surfaces require a Claude subscription or a Console account.
Homebrew and WinGet installs do not auto-update; only the native install does.
It is an agent with real command access - the security model is whatever permissions and hooks you configure, not the tool's defaults.

benchr has not run this capability. No test has been published. 1 self-reported run awaiting independent review.

benchr test runs →

Task success rate on any specific repository or language
How results compare with other coding agents - no neutral benchmark is published

Source-check history

September 1, 2026DocsOfficially documentedFirst logged from the current product overview, including the four surfaces and the install channels that do not auto-update.