Workbench

English | 简体中文 | 日本語

See the port. Find the process. Know the cost.

A local-first desktop workbench for people who run services and AI coding agents on their own machine. One window shows every process, listening port, dev server, project, task — and what your Claude Code / Codex sessions actually cost. Data never leaves the device, and there is no account, no cloud, no model to configure.

What makes it different

Port → process → cost	A listening port links to the process that owns it, and a process shows every port it holds. Session files connect the same machine's work to real spend. No other local tool joins those three views.
Understands dev servers	Not just "port open": Vite, Next, Nuxt, Django, uvicorn, Rails and 30+ more are recognised from the command line, with a confidence level.
Keyboard-first	Ctrl/Cmd+K searches every page, process, port, project and task. Flat table or parent/child tree. Everything reachable without the mouse.
Trustworthy by construction	Protected system processes are rejected in Rust, the target identity is re-checked right before a kill, UAC elevation is opt-in per action, and every report states what it verified.
Local-first, honestly	SQLite ledger on your disk, CSP locked to self, no telemetry. The only outbound request is the optional probe of a local opencode server.

Screenshots

Overview	Processes
CPU / memory trends, disks, process groups	Full snapshot, virtual scrolling, CSV export

Process tree	Command palette
Parent/child grouping with per-process ports	Ctrl/Cmd+K across pages, processes, ports, projects, tasks

Servers & ports	AI sessions
netstat / ss parsing, dev-framework tags	Session files, live sessions, tool mix

Cost ledger	Tasks
Budget progress, trend, model split	Drag-and-drop kanban

30-second tour

Recorded from the real UI in browser demo mode (no real host, path or account data appears in it). Download the MP4.

Features

Feature	What you get
Monitoring	Live processes, listening ports (netstat on Windows, ss -tulpn on Linux with a netstat fallback) and dev-server detection with framework + confidence
Process tree	Flat table or parent/child tree from parent_pid; orphan and cycle-safe, virtual scrolling, per-row detail with the process's own listening ports, CSV export
Command palette	Ctrl/Cmd+K fuzzy search over pages, processes, ports, projects and tasks; keyboard navigation with ↑ ↓ ↵ Esc
Elevated kill	"End as administrator" retry after a normal kill fails — triggers the OS UAC prompt; protected system processes are rejected by the backend before any prompt appears
AI cost metering	Token costs of local Claude / Codex sessions in a SQLite ledger — no model, no API key, nothing leaves your machine
Pricing editor	40+ built-in model prices; edit / add / reset overrides from the UI
Backup & restore	Timestamped snapshots with one-click rollback and integrity checks (VACUUM INTO + PRAGMA integrity_check)
Start at login	Real OS registration (registry / autostart entry), toggle reads back from the OS
Environment check	Per-item ok / warn / fail report for platform, data directory, cost DB, session folders and port tools, with a copyable summary
CLI mode	The same binary answers workbench ps, workbench ports and workbench diag (add --json) — read-only, scriptable, no window
Visual system	7 overview layouts × 8 themes × 3 shells, draggable & hideable widgets with per-widget widths, theme editor with share links
Trilingual UI	简体中文 (default) / English / 日本語
Cross-platform	Windows (NSIS + portable exe), Linux (.deb), macOS (.app, CI-built unsigned)

Download

v0.3.0 — grab the installer for your platform:

Platform	File
Windows 10/11 x64	Workbench_0.3.0_x64-setup.exe (NSIS, ~5 MB) — attached to the GitHub / Gitee / GitCode release
Linux x64	no published asset yet — npm run tauri build -- --bundles deb, or take the linux-deb CI artifact
macOS	no published asset yet — npm run tauri build -- --bundles app, or take the macos-app CI artifact (unsigned: right-click → Open)

Not code-signed, so Windows SmartScreen may show an "Unknown publisher" warning — choose More info → Run anyway. macOS builds are unsigned too; right-click the app and choose Open the first time. Certificates cost money; neither file is unsafe.

Pick any mirror — identical code, each publishes its own releases:

Mirror	Releases
GitHub	github.com/X33834/workbench/releases
Gitee	gitee.com/badhope/workbench/releases
GitCode	gitcode.com/badhope/workbench/releases

Build

Requires Node.js ≥ 20 and Rust ≥ 1.90:

npm install
npm run tauri dev      # desktop dev mode (hot reload)
npm run tauri build    # installers → src-tauri/target/release/bundle/

Local CI mirror (same steps as .github/workflows/ci.yml):

node scripts/ci-local.mjs          # typecheck → lint → format → unit → build → Playwright → cargo
node scripts/ci-local.mjs --fast   # skip the visual-regression pass

Command line (same executable, no window):

workbench ps            # every process: pid, name, cpu%, memory, parent, status
workbench ports         # listening TCP/UDP ports with the owning PID
workbench diag          # environment self-check
workbench ports --json  # machine-readable output for scripts

Tech stack, project structure, FAQ, roadmap, screenshots and everything else live in the architecture & reference doc (简体中文 / 日本語). How this compares with Mission Center, System Informer, btop++, Glances, ccusage and friends: docs/COMPARISON.md.

Contributing & License

• Contribution workflow: CONTRIBUTING.md · Security: SECURITY.md · Conduct: CODE_OF_CONDUCT.md · Changelog: CHANGELOG.md
• MIT © 2026 x1882