CC Mode Switcher

Multi-role scheduler for Claude Code. Define any number of roles (default: 🧠 Planner + ⚙️ Worker), bind a different model to each — e.g. a reasoning model for planning and a fast model for execution — then launch Claude Code in the right role from a built-in multi-tab terminal (xterm.js + node-pty) or your favourite external terminal. Tab = one role, one pty session, one tool allow / deny list.

Built with Electron + Vue 3 + TypeScript. Disk-backed YAML configuration (no localStorage for the things that matter).

Design Philosophy

The idea comes from how Anthropic routes work internally: planning-shaped tasks go to certain models, execution-shaped tasks to others. Think of a company — the cognition you expect in leadership is different from what you need in the execution layer; the people who make the plan and the people who carry it out are never the same group, and their cost is never the same either.

This tool can't be that smart. It draws the simplest possible line — plan vs. everything else — and, because it sits on top of Claude Code, leaves the routing decision to a human convention: you decide which model thinks (Plan) and which model executes (Work). Expensive reasoning tokens are spent only where they matter; cheap models do the mechanical work.

Features

🤖 Model Management

Add / edit / duplicate / delete model configs (display name, base URL, API key, model ID)
Provider presets with URL autocomplete — GLM (Zhipu), MiniMax, DeepSeek, Kimi (Moonshot), Z.ai, Qwen (DashScope)
Type a name or model ID (glm-5.3, MiniMax-M3, …) and the base URL auto-fills
Quick-pick model ID chips once a provider is matched
Connection test per model — green toast with latency (connected in 143ms) or red toast with the network error

🎭 Multi-role Scheduler (Rancher-for-roles)

Any number of roles, defined in ~/.cc-mode-switcher/roles.yaml — no hard-coded list of roles in the app
Each role picks: a model, an extended-thinking toggle, a system-prompt file path, allowed tools, denied tools, and denied plugins
Default roles are 🧠 Planner (read-only) and ⚙️ Worker (writes files) — built-in prompts at ~/.cc-mode-switcher/prompts/{plan,worker}.md define the .cc-delivery/ file contract between roles
Table view for clicks: filter, inline cell edits, right-click to duplicate / delete
YAML view for power users: live syntax check, dirty-state indicator, save reloads the table
Two-way same source: editing the table writes YAML; editing YAML and saving refreshes the table
Reset restores the default plan + worker roles but keeps your models and any edits you made to the prompt files

⌨️ Built-in Multi-tab Terminal

xterm.js + node-pty in the main process — each tab is one pty session, bound to one role + one model
Tab title: {project} | {roleLabel}({modelName}) (e.g. acme-web | 🧠 Planner(glm-5.3))
Right-click a tab to clone (reuse the same role + cwd) or detach to a separate window
Mac shortcuts (active only when the terminal pane has focus):
- Cmd+T — clone the focused tab
- Cmd+N — open the role picker and start a new session
- Option+T — start the role currently selected on the left
Detached window exposes a single "↩ merge into main" button that transfers the session back
Per-session snapshot: changing role bindings afterwards does not affect already-open tabs

🚫 Zero-touch on Claude Code settings

Claude Code's ~/.claude/settings.json env block would otherwise override terminal env vars. We sidestep it with two flags per session:

--setting-sources "" — Claude Code skips all default settings files (user / project / local)
--settings "$CC_MODE_DIR/<ModelName>.json" — loads the per-session temp JSON (named after the bound model) at highest priority, overriding everything
The app never reads or writes ~/.claude/settings.json or any settings file — no backups, no surprises

📦 Release & Versioning (GitHub Actions, manual)

macOS / Windows / Linux installers are published to GitHub Releases — download the latest build here, or see the download page for per-OS install notes.

Releases are driven by three manual workflows — no automatic triggers, no local CLI needed:

List releases — see what already exists before doing anything
Set version & tag — bump up (patch / minor / major), downgrade, or set any version. Writes a release vX.Y.Z commit and pushes the tag
Release Electron App — builds the three OS targets in parallel and creates / updates the GitHub Release

Full operational guide: Release & Versioning Workflow.

⚙️ Settings & Polish

🌙 Dark (default) / ☀️ Light theme — CSS-variable based, toggle in the toolbar
English (default) / 简体中文 — quick toggle in the toolbar
Icon-only action buttons with hover tooltips
Centered toast notifications
Models + role config persists in ~/.cc-mode-switcher/*.yaml; UI preferences (theme, language) in localStorage

Generated Launch Scripts

The detail panel below the role table shows the full shell snippet for the selected role — same content the built-in terminal runs and same what the "open in external terminal" button sends. Example, plan role bound to GLM-5.3:

# cc-mode-switcher · plan
export ANTHROPIC_BASE_URL="https://open.bigmodel.cn/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-..."
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.3"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.3"
export ANTHROPIC_DEFAULT_FABLE_MODEL="glm-5.3"
export ANTHROPIC_DEFAULT_OPUS_MODEL_NAME="glm-5.3"
export ANTHROPIC_DEFAULT_SONNET_MODEL_NAME="glm-5.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME="glm-5.3"
export ANTHROPIC_DEFAULT_FABLE_MODEL_NAME="glm-5.3"
export ANTHROPIC_MODEL="glm-5.3"
export CLAUDE_CODE_SUBAGENT_MODEL="glm-5.3"
export MAX_THINKING_TOKENS=16000

# Per-role temp settings file (priority > ~/.claude/settings.json)
CC_MODE_DIR=$(mktemp -d -t cc-mode-XXXXXX)

cat > "$CC_MODE_DIR/GLM-5.3.json" <<'CCMODE_EOF'
{
  "env": { ... same 13 keys as above ... }
}
CCMODE_EOF

# Launch claude with role-scoped prompt + tool permissions
exec claude --setting-sources '' --settings "$CC_MODE_DIR/GLM-5.3.json" \
            --system-prompt-file ~/.cc-mode-switcher/prompts/plan.md \
            --disallowed-plugins 'superpowers' \
            --allowedTools Read,LS,Glob,Grep \
            --disallowedTools Edit,Write,NotebookEdit,Bash

superpowers is always forced into --disallowed-plugins regardless of role config, and every session uses --setting-sources "" so ~/.claude/settings.json is never read.

Selecting the Work card instead regenerates the same block with Work's bindings — MAX_THINKING_TOKENS line is omitted, the temp file becomes e.g. $CC_MODE_DIR/MiniMax-M3.json, and the alias is cc-w='claude --setting-sources "" --settings "$CC_MODE_DIR/MiniMax-M3.json"'. Switch cards any time and click ▶️ again to open a fresh terminal in the new mode.

--setting-sources "" skips every default settings file (user / project / local), so ~/.claude/settings.json never loads. --settings "$CC_MODE_DIR/<ModelName>.json" then loads the per-mode temp JSON at higher priority than any other source, making it the single source of truth.

Requirements

Dependency	Version	Notes
Node.js	≥ 22 (24+ recommended)	Vite 7 requires Node 20.19+ / 22.12+; this project targets Node 24
pnpm	≥ 9	Preferred — the lockfile is `pnpm-lock.yaml` (npm ≥ 10 also works)
macOS	12+	Apple Silicon or Intel. Packaging installers needs Xcode Command Line Tools (`xcode-select --install`)

Windows/Linux: the UI builds and runs, but "Open in Terminal" is macOS-only for now (it uses AppleScript / .command files).

⚠️ pnpm ≥ 10 note

pnpm 10 blocks dependency postinstall scripts by default, so Electron's binary never downloads and npm run dev fails with Error: Electron uninstall. Fix:

pnpm approve-builds   # select electron (and esbuild)

or add to package.json before installing:

"pnpm": { "onlyBuiltDependencies": ["electron", "esbuild"] }

Mirrors for mainland China

Speed up dependency and Electron binary downloads:

# npm/pnpm registry
pnpm config set registry https://registry.npmmirror.com

# Electron binary + electron-builder helper binaries
export ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/
export ELECTRON_BUILDER_BINARIES_MIRROR=https://npmmirror.com/mirrors/electron-builder-binaries/

Or commit a project-level .npmrc:

registry=https://registry.npmmirror.com
electron_mirror=https://npmmirror.com/mirrors/electron/
electron_builder_binaries_mirror=https://npmmirror.com/mirrors/electron-builder-binaries/

Getting Started

pnpm install   # or npm install
npm run dev    # start dev server + Electron

Other scripts:

npm run build   # build main/preload/renderer bundles
npm run dist    # package installers (electron-builder)

Extending Provider Presets

Add an entry to src/renderer/src/data/providers.ts — the form autocomplete, keyword auto-fill, and model ID chips pick it up automatically:

{
  id: 'glm',
  name: 'GLM (Zhipu)',
  baseUrl: 'https://open.bigmodel.cn/api/anthropic',
  keywords: ['glm', 'bigmodel', 'zhipu', '智谱'],
  models: ['glm-5.3', 'glm-5.3-air', 'glm-4.6', 'glm-4.5']
}

Any Anthropic-compatible endpoint works.

Documentation

Full chaptered guides live in docs/ — English by default, 简体中文版 available:

Quick start — first-time setup in 5 steps
Models & providers — presets, env vars, override guard, model pairing
Plan mode playbook — produce plan documents as the intermediate artifact
Work mode playbook — execute the plan strictly, never re-plan mid-flight
End-to-end example — one feature from requirement to delivery
Release & versioning workflow — GitHub Actions workflows for cloud builds, version up / down, GitHub Releases (all manual, no local CLI)
Local build — clean install: clear node_modules, pnpm store, electron / electron-builder caches, then reinstall (when pnpm run dev / pnpm run dist misbehaves)

Project Structure

src/
├── main/            # Electron main process (IPC, terminal launch)
├── preload/         # contextBridge API
└── renderer/
    └── src/
        ├── components/    # Vue SFCs (panels, cards, modal, toast, …)
        ├── composables/   # models, i18n, theme, terminal, toast
        ├── i18n/          # en / zh message catalogs
        ├── data/          # provider presets
        └── assets/        # global styles

Release Workflow

Publishing is a two-step manual process — nothing fires automatically. Both steps happen entirely on GitHub via the Actions tab; no local CLI required.

Step	Workflow	What it does
1	Set version & tag	Writes a `release vX.Y.Z` commit, pushes the `vX.Y.Z` tag to origin. Old tags/releases are untouched.
2	Release Electron App	Builds mac / win / linux artifacts and creates/updates the GitHub Release.

A third helper, List releases, prints what's already on the server so you can decide whether to bump, downgrade, or re-publish.

First-time setup (one-off, on GitHub)

Repo → Settings → Actions → General → Workflow permissions → Read and write permissions → Save. Without this, the runner can't push back to the repo.

See what already exists

Actions → List releases → Run workflow → wait → open the run → expand Print releases + tags. You get two lists:

Releases (via gh release list) — every published GitHub Release with status (Published / Draft / Pre-release)
Tags (via git ls-remote --tags) — every tag, including ones whose Release was deleted

If a tag shows up in the second list but not the first, its Release was deleted — re-publish it via step 2 with that tag.

Bump up (auto, patch / minor / major)

Actions → Set version & tag → Run workflow:

Input	Value
`mode`	`auto`
`bump`	`patch` (or `minor` / `major`)
`version`	(leave blank)

What happens: bumps package.json + pnpm-lock.yaml + the version label in SettingsPanel.vue, commits release vX.Y.Z, pushes the new tag. Nothing is built yet — go to step 2.

Set to an explicit version (upgrade OR downgrade)

Actions → Set version & tag → Run workflow:

Input	Value
`mode`	`set`
`bump`	(ignored)
`version`	`2.0.0` (or anything — lower than current is a downgrade, e.g. `0.9.6`)

What happens: same as above, but the target version is whatever you typed. Downgrades are non-destructive — the previous tag and its Release stay in place.

If the tag you typed already exists on origin, the workflow aborts and tells you to either pick a different version or re-publish via step 2.

Build & publish the GitHub Release

Actions → Release Electron App → Run workflow:

Input	Value
`tag`	(leave blank to build whatever is currently on main — i.e. the commit step 1 just pushed)

The runner checks out that ref, runs electron-builder --publish always, and creates a GitHub Release named after the version in package.json. mac / win / linux all build in parallel.

Re-publish an existing tag (accidental delete, rebuild, …)

Actions → Release Electron App → Run workflow:

Input	Value
`tag`	`v1.0.0` (fill the tag you want to re-publish)

The runner checks out that tag, rebuilds, and overwrites the existing Release with fresh artifacts.

Deleting releases / tags

The workflows never delete. To clean up, do it from GitHub's web UI (repo → Releases → trash icon on the release) or from any terminal that has the gh CLI:

# Delete just the Release (keep the tag — re-publishable via step above)
gh release delete v1.0.0

# Delete Release + tag
gh release delete v1.0.0 --yes
git push origin --delete v1.0.0

Or mark it Draft in the web UI to hide without losing it.

FAQ

Why doesn't this app touch my ~/.claude/settings.json? Because it doesn't need to. Every alias launches Claude Code with two flags: --setting-sources "" skips user / project / local settings files entirely, and --settings "$CC_MODE_DIR/<ModelName>.json" loads a per-mode temp JSON (named after the bound model, regenerated on every click) at higher priority than any source — including ~/.claude/settings.json. App settings live in $CC_MODE_DIR, never in your home directory.

The TSM AdjustCapsLock… / IMKCFRunLoopWakeUpReliable log lines in dev? Harmless macOS input-method noise present in every Electron app. The window auto-reloads if the renderer ever crashes.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
_bak/old_ver		_bak/old_ver
build		build
docs		docs
scripts		scripts
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
README.zh-CN.md		README.zh-CN.md
electron.vite.config.ts		electron.vite.config.ts
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
tsconfig.json		tsconfig.json
tsconfig.node.json		tsconfig.node.json

Navigation Menu

Repository files navigation

CC Mode Switcher

Design Philosophy

Features

🤖 Model Management

🎭 Multi-role Scheduler (Rancher-for-roles)

⌨️ Built-in Multi-tab Terminal

🚫 Zero-touch on Claude Code settings

📦 Release & Versioning (GitHub Actions, manual)

⚙️ Settings & Polish

Generated Launch Scripts

Requirements

⚠️ pnpm ≥ 10 note

Mirrors for mainland China

Getting Started

Extending Provider Presets

Documentation

Project Structure

Release Workflow

First-time setup (one-off, on GitHub)

See what already exists

Bump up (auto, patch / minor / major)

Set to an explicit version (upgrade OR downgrade)

Build & publish the GitHub Release

Re-publish an existing tag (accidental delete, rebuild, …)

Deleting releases / tags

FAQ

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

CC Mode Switcher

Design Philosophy

Features

🤖 Model Management

🎭 Multi-role Scheduler (Rancher-for-roles)

⌨️ Built-in Multi-tab Terminal

🚫 Zero-touch on Claude Code settings

📦 Release & Versioning (GitHub Actions, manual)

⚙️ Settings & Polish

Generated Launch Scripts

Requirements

⚠️ pnpm ≥ 10 note

Mirrors for mainland China

Getting Started

Extending Provider Presets

Documentation

Project Structure

Release Workflow

First-time setup (one-off, on GitHub)

See what already exists

Bump up (auto, patch / minor / major)

Set to an explicit version (upgrade OR downgrade)

Build & publish the GitHub Release

Re-publish an existing tag (accidental delete, rebuild, …)

Deleting releases / tags

FAQ

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages