claudeworks

ClaudeWorks Run separate Claude Code setups without hand-editing config.

Home  ·  Docs  ·  Releases  ·  Issues

Each profile is an isolated Claude Code config directory: its own MCP servers, skills, agents, slash commands, and settings. Work in one terminal, personal in another, both running Claude Code with different tools loaded. 647 curated plugins and 4,795 skills come built in.

---

Why this exists

I kept managing two Claude Code setups by hand, the work one with its MCP servers and skill set, the personal one with a different CLAUDE.md and different tools. Swapping meant editing ~/.claude.json and mcp.json by hand, then undoing it next session. Claude Code has no native multi-profile support, so I built this instead.

Plugin caches are shared across profiles via symlinks, so you install a plugin once and updates propagate everywhere.

What it does

Each profile runs in its own CLAUDE_CONFIG_DIR, which is Claude Code's built-in way to redirect its config root. MCP servers, skills, agents, and slash commands from profile A can't appear in profile B's session. Real per-process isolation, not config swapping.

A curated plugin marketplace ships inside the app. Browse 114 upstream marketplaces and 647 featured plugins without leaving ClaudeWorks. One search bar queries 7,009 flat entries (every marketplace, plugin, skill, command, agent, and MCP server) at once.

Every profile ships arbitrary shell aliases. Each one cd's to a directory and optionally auto-runs /workflow or a saved prompt. One profile can claim the bare claude command, so plain claude in any terminal lands in your default profile without the bloat of the others.

ship-api     → cd ~/code/api,    then /workflow
debug-mobile → cd ~/code/mobile, then runs a prefilled prompt
claude       → opens Default profile (bare-claude interception)

More

Per-profile /workflow commands (and named variants like /workflow-debug, /workflow-deploy) become real Claude Code commands in that profile's session. Variants can be scoped to specific launch directories.

Multi-agent Teams (experimental, behind CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1): compose a team from existing profiles with drag-and-drop, assign roles, pick a lead, and launch. The lead opens with a generated /start-team that spawns each teammate.

Per-directory MCP management: disable individual servers inside sensitive repos while keeping them available elsewhere.

Status bar builder: 17 widgets, drag to reorder, named save/load, per-profile override.

Profile export / import: a self-contained JSON that lists exactly which plugins the target machine is missing.

Profiles Doctor: read-only health check across profiles, plugins, and alias scripts. Repair mode is opt-in and always writes a .bak-<ts> first.

Local skill provenance: skills in ~/.claude/skills/ show a skillfish or git origin tag when the source is detectable.

What it doesn't do

macOS only at launch. Windows is planned and Linux likely follows. The Electron shell is already cross-platform, but the terminal launch and Keychain sync need ports.

Profiles live on your machine, no cloud sync. Use exportProfile / importProfile to move them.

No GitHub token required. ClaudeWorks tries the gh CLI first, then GITHUB_TOKEN, then falls back to anonymous fetches (60 req/h). See Architecture for the fallback chain.

Teams need Anthropic's experimental flag (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1). The feature works today but expect behaviour to shift as the upstream matures.

We don't generate or manage Claude API keys. Run claude login first. The app copies credentials from your existing login.

Install

From a release

1. Download the latest .dmg from the Releases page.
2. Open the DMG and drag ClaudeWorks to Applications.
3. First launch, macOS will flag the app as unsigned; right-click, Open, confirm.
4. Run gh auth login if you haven't. ClaudeWorks uses the gh CLI for GitHub API calls, which gives you 5,000 req/h instead of 60.

From source

git clone https://github.com/Mduffy37/claudeworks.git
cd claudeworks
npm install
npm run build
npm start

Requires Node 20+, Electron 33, and the gh CLI on your PATH for marketplace features. Run npm run dev for the Vite dev server + Electron watch loop.

Quickstart

Create your first profile

1. Launch ClaudeWorks. The sidebar shows your profiles (the app seeds a profile-creator workspace to help you author one).
2. Click New profile in the sidebar dock.
3. Pick which plugins, skills, agents, and MCP servers load in this profile. Everything defaults to off. Opt in to what you want.
4. Click Save.

Launch a profile

Click the Launch button on the profile card. A terminal opens with CLAUDE_CONFIG_DIR already set. Terminal.app is the default; iTerm2 can be selected in Settings, with more terminal options on the way.

Or type your shell alias. If you added aliases: [{ name: "ship-api", directory: "~/code/api" }] to the profile, ship-api from any terminal cd's to that directory and starts a Claude Code session in that profile.

Discover plugins

Open Configure Claude → Plugins → Browse. The global search bar queries every marketplace, plugin, skill, command, agent, and MCP server at once (7,009 entries across 114 marketplaces). Click any plugin card for the detail pane: upstream README rendered inline, with a peer-plugin list beside it.

Architecture

Profiles as config directories

Each profile lives at ~/.claudeworks/profiles/<name>/config/ and launches with CLAUDE_CONFIG_DIR pointing at it. This is Claude Code's built-in way to redirect its config root per session, so each profile's MCP servers and skills live in their own namespace.

Plugin caches under ~/.claude/plugins/cache/ are shared across profiles via