Yardsort
Run coding agents side by side in separate Git worktrees, then review and carry their work forward.
- Source repo
- joaoh82/yardsort
- Stars
- ★ 10
- Last updated
- 2d ago
- License
- GPL-3.0
- Primary language
- Rust
- FA score
- 70/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer exploring several implementation approaches can give each agent a separate worktree in the same repository.
- Not a fit if
- Organizations that require signed Windows installers
- Teams expecting the app to bundle coding agents
- Source review
- 70/100 · Some gaps
What does this agent do, and when should you use it?
Yardsort is an open-source desktop app for Linux, macOS, and Windows, built with Tauri and Rust. For each task, it creates a Git branch and worktree, then launches the coding agent you choose in a real terminal. You can run several workspaces in parallel and inspect their terminals, changed files, and diffs in the app. Its companion `ys` CLI can start workspaces without opening the window, while a background terminal process keeps agents running after the window closes. The results remain ordinary Git branches and worktrees; you install and sign in to agents yourself, and GitHub issue and pull request features use your logged-in `gh` CLI.
After you add or clone a Git project, describe a task and choose a configured harness; Yardsort creates a branch and Git worktree, then starts the agent through a real PTY. The app displays the terminal, changed files, file tree, and diffs, and can edit text files while retaining drafts and checking for concurrent disk changes. You can resume or fork sessions, prepare a handoff message from the local activity record, and commit, push, or open a pull request. The ys CLI can create workspaces, attach to terminals, print logs, generate handoff messages, and run workflows. When GitHub features are used, Yardsort reads and manages issues and pull requests through GitHub CLI. Yardsort does not read agent output or credentials; optional Assist sends specified diffs and task text to Jev only when enabled with the user's TypeSafe API key.
- A developer exploring several implementation approaches can give each agent a separate worktree in the same repository.
- A maintainer who wants to review and ship agent changes from one desktop app can inspect diffs, then commit, push, or open a pull request.
- A team using terminal tools such as Claude Code, Codex, or OpenCode can configure them as harnesses and switch among them within a project.
- A maintainer triaging GitHub issues can open Tasks and start a workspace with the issue context as the agent's first message.
- A developer who wants to script coding tasks from a terminal can use
ysto create workspaces, attach to sessions, or run workflows.
How do you install or deploy this agent?
Download a release from GitHub: Linux offers AppImage, deb, and rpm packages; macOS offers a universal DMG or Homebrew installation; Windows offers a setup executable or MSI. You also need Git and at least one installed, signed-in agent CLI; agents are not bundled. GitHub issue and pull request features require the GitHub CLI to be installed and logged in. In-app updates are documented for macOS, Windows, and the Linux AppImage; the README does not specify in-app updates for other Linux packages.
How do you use this agent?
Open Yardsort and open a local Git repository or clone one. Create a workspace from the project, enter a task, select an agent, and press Enter. Follow the agent in the terminal and review its files and diffs in the app; when ready, commit, push, or open a pull request. You can also start a workspace from a terminal:
ys workspace new <project> "<prompt>"The CLI hands the agent to the background process; use ys attach to return to its terminal session and ys logs to inspect a session that has ended.
What are this agent's strengths and limitations?
- Each task gets its own Git worktree and branch, so agents can work in parallel without changing the main checkout.
- Linux, macOS, and Windows are first-class platforms, with the README stating that all three are built and tested on each change.
- It starts several terminal agents and lets you configure others in
settings.tomlwithout waiting for an app release. - A background terminal process keeps work running after the window closes; sessions can be resumed, forked, or handed off using the local activity record.
- GitHub issues and pull requests use the user's own GitHub CLI, so the app does not hold GitHub credentials.
- The app does not bundle coding agents; users must install and sign in to at least one compatible agent CLI.
- Windows installers are not yet code-signed, so users may need to choose the operating system's run-anyway option.
- GitHub issue and pull request features depend on a logged-in GitHub CLI; Assist separately requires a TypeSafe API key and sends selected content.
- The project describes itself as early and moving fast, and says users should expect rough edges.
How does this agent compare with similar options?
The README compares Yardsort with Conductor and Superset, highlighting first-class Linux, macOS, and Windows support as a distinction.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Yardsort This agent | 70 · Some gaps | Desktop appFree | ★ 10 | 2d ago | Rust | Codex · Claude Code |
| Helmor Local Agent Workbench | 52 · Major gaps | Desktop appFree + model costs | ★ 1.3k | 2mo ago | TypeScript | Codex · Claude Code |
| Runner Multi-Agent Terminal | 82 · Some gaps | Desktop appFree | ★ 275 | today | Rust | Codex · Claude Code |
| Intent Agent Workbench | 60 · Some gaps | Desktop app | ★ 305 | today | JavaScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The security notes say the app launches user-configured programs with an argument list rather than through a shell. Those programs inherit the login-shell environment, and an agent can act within its own permission model, so least privilege is only partly supported. The README and SECURITY.md repeatedly describe confirmation before actions such as closing with work in flight, deletion, merging, closing, or reopening; updates also require user action. Network uses, the update check, optional Assist, and local log scope are described concretely. The app says it stores no credentials, but it uses the user's GitHub CLI login and can use a user-provided Assist key. Dependency security has a frozen-lockfile install process but no evidence of dependency auditing or vulnerability handling. External effects have confirmations and permission explanations, though some actions reach GitHub and other services. Ordinary git worktrees, retained branches, and confirmation before discarding uncommitted changes provide some recovery options; the files do not give a complete restoration procedure. Vulnerability reporting and release notes provide attribution paths, but publisher identity is unverified and maintainer identity cannot be confirmed.
The README and Cargo and package metadata agree on version 0.20.0. CI covers Linux, macOS, and Windows and lists formatting, lint, type-check, and test steps; this supports ordinary consistency, but this review did not execute verification. GitHub CLI and terminal agents are optional or external dependencies. The README documents prerequisites for some features but cannot guarantee those external tools remain available. CI and packaging workflows give explicit messages for timeouts, missing credentials, and some failures; evidence for runtime error handling is limited.
The intended audience, parallel worktree model, supported platforms, built-in agents, CLI, tasks, and workflows are described concretely for developers using terminal coding agents. The security notes state that the app launches configured programs and that agent capabilities depend on their permission model, but the broad feature descriptions do not specify permission boundaries or risks for every feature. Workflows describe triggers and variable validation, and the README lists controls, commands, and prerequisites; the supplied material does not corroborate every operation's trigger conditions. CI has a Linux, macOS, and Windows matrix, providing strong platform-fit evidence.
The README is organized by feature and links to a quick start and topic guides; however, the supplied README truncates in the supported-agents section, limiting assessment of documentation coverage. A quick-start link and download entry are present, but the supplied files do not provide installation steps. Yardsort, ys, and workspace terminology is consistent in the examples, and Cargo and package metadata agree on the version; no changelog evidence is included. Examples cover CLI use, workflows, handoff content, and supported agents; FAQ evidence is limited. The docs identify limits such as log coverage, cost-estimate scope, the GitHub CLI requirement, and security fixes being limited to the latest release. GPL-3.0-only metadata agrees with the GPL v3 LICENSE. Security reporting and release notes provide maintenance paths, but publisher identity is unverified and no specific maintainer responsibility is established.
The product describes outputs including separate worktrees, live file diffs, activity records, handoff messages, and optional workflows that support review and follow-up; the supplied material cannot verify runtime results. It combines parallel terminal agents with ordinary Git workspaces, desktop review, and a CLI, showing a clear additional use. Cost notes distinguish estimated API prices from subscription bills and state that some models have no cost estimate; actual resource use and cost effectiveness cannot be established statically.
The README and SECURITY.md make specific claims about data flows, security boundaries, confirmations, and limitations. Cargo, package metadata, the license file, and CI allow some claims about version, license, and declared checks to be cross-checked. Multiple files corroborate license, platform, and version information, but most product behavior is still described in the README and security notes. The material distinguishes recorded metadata, unknowns, and estimates—for example, handoff content says it does not know the agent conversations—but many features and security guarantees are not demonstrated by implementation files in the supplied evidence, so only partial credit is justified.
- The security notes state that terminal programs inherit the login-shell environment and agent permissions depend on the agent's own permission model; do not treat worktree isolation as operating-system permission isolation.
- GitHub, terminal agents, and Assist involve user logins, user-provided keys, or third-party services. Check what will be sent and note that unknown models are excluded from API cost estimates.
- This assessment uses only the supplied static files; the app and tests were not run. The README also truncates in the supported-agents section, so behavior, installation details, and maintainer identity cannot all be verified.
FAQ
Do I need a paid subscription to use Yardsort?
Does Yardsort upload my code or agent credentials?
Can I use Yardsort without installing an agent?
Will closing the app window stop an agent?
What is required for GitHub issues and pull requests?
gh). Yardsort uses it to access and manage GitHub content with your existing permissions.