Claude Code doesn't ship with a control panel. This is it.
You're running Claude Code in VS Code, but you can't see what MCP servers are active, which skills are loaded, whether your team's config matches yours, or what your agents are actually doing. Setup is scattered across dotfiles, JSON configs, and tribal knowledge.
Claude Code ToolBox gives you:
- Agentic Teams — 10 specialized AI agents (architect, security, frontend, backend, QA...) that debate designs, plan implementations you approve, then execute together. One Claude session becomes a whole engineering team.
- A visual MCP & skills hub — see, search, install, and manage everything from one sidebar
- Live Agent Dashboard — every running Claude Code session on your machine, with real-time cost tracking and context visibility
- One-click migration — bring your Cursor or Copilot setup (MCP, rules, skills) into Claude Code automatically
- AntiVibe Safety Guards — block destructive commands (rm -rf, git push --force, DROP TABLE), enforce domain whitelisting for web requests, and prevent supply chain attacks (blocks known-compromised packages like event-stream, node-ipc, colors). Configurable patterns, allow-overrides, block/warn modes
- Token Optimization — reduce token usage 30-60% with project maps, read deduplication, output compression, .claudeignore, verbosity control, and context budget alerts
- Workspace-aware context priming — Claude Code finally knows what your project actually has configured
2,000+ installs · Works on macOS, Windows, Linux · Also ships as a JetBrains plugin
Install from VS Code Marketplace · GitHub
Agentic Teams: Why One Agent Isn't Enough
A single Claude Code session does everything: plan, code, review, test. That's like one person doing every role. Agentic Teams splits the work across specialized agents that challenge each other — the architect plans, the security reviewer pokes holes, the QA engineer writes tests, and you approve before anyone writes code.
10 role-based agents collaborate through 8 protocols:
- Debate + Judge — agents argue both sides, a judge synthesizes
- Plan-then-Code — architect plans, you approve, devs execute
- Converge — parallel thinking → cross-pollination → synthesis
- Swarm dispatch — every team is a
/command that fires all agents in parallel
One-click setup enables CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS and installs agents to ~/.claude/agents/ where Claude Code discovers them natively.
Skill-Backed Agents with Long-Term Memory
Each agent can be pointed at a SKILL.md instead of a freeform prompt — structured, reusable expertise. Agents also have per-agent long-term memory: they learn from interactions and retain knowledge across runs.
Agent Dashboard: Live Visibility
Opt-in dashboard shows a card for every running Claude Code session. Token-based cost estimates, context-window fill %, live tool feed, grouped by workspace.
AntiVibe Safety Guards: Block Destructive Commands + Domain Whitelisting + Supply Chain Protection
Claude Code runs arbitrary shell commands, fetches URLs, and installs packages. AntiVibe Safety Guards installs Claude Code hooks that prevent catastrophic mistakes, data exfiltration, and supply chain attacks:
| Guard |
What it protects against |
| Destructive Command Block |
rm -rf, git push --force, git reset --hard, git branch -D, DROP TABLE, curl \| sh, chmod 777, and 30+ more patterns |
| Domain Whitelisting |
Only allows web requests to known-good domains (GitHub, npm, docs sites, etc.) — blocks exfiltration to pastebins, file-sharing, or unknown hosts |
| Supply Chain Guard |
Blocks installation of known-compromised packages (event-stream, node-ipc, colors, faker, ua-parser-js, peacenotwar, es5-ext) — prevents npm/pip/yarn/pnpm/gem/cargo supply chain attacks |
Triple-Confirmation Flow (Block mode)
All three guards use a triple-confirmation pattern in block mode:
| Attempt |
What happens |
| 1st |
Blocked — Claude receives a warning about the dangerous operation |
| 2nd |
Blocked — hook instructs Claude to ask the user for permission |
| 3rd |
Allowed — only proceeds because the user explicitly confirmed |
The counter resets after each allow cycle (next occurrence requires fresh confirmation). State is session-scoped and expires after 1 hour.
Modes
- Block (default) — triple-confirmation as above; Claude must ask you before proceeding
- Warn — hook emits stderr warning but allows execution (for teams who want visibility without hard stops)
- Allowlist (domains, default) — only listed domains pass; everything else triggers triple-confirmation
- Blocklist (domains) — all domains allowed except explicitly blocked ones
Customization
- Allow overrides — permit specific patterns despite the blocklist (e.g.,
rm -rf node_modules if your workflow needs it)
- Add your internal domains —
*.internal.company.com in the allowlist
- Edit blocked packages — add or remove packages from the supply chain blocklist in Settings
- Edit directly — use the command palette or hub quick-actions to view/edit patterns and domains
One Click Setup Integration
Token Optimization is enabled by default when you run One Click Setup. AntiVibe Safety Guards must be enabled separately via the hub toggle or command palette. After enabling, you'll be prompted to close and reopen VS Code for hooks to take effect.
Commands
| Command |
What it does |
AntiVibe Safety Guards — Enable |
Installs hooks to ~/.claude/settings.json |
AntiVibe Safety Guards — Disable |
Removes hooks and scripts |
AntiVibe Safety Guards — Status |
Shows active patterns, modes, and hook state |
AntiVibe Safety Guards — Edit destructive patterns |
View and customize blocked patterns |
AntiVibe Safety Guards — Edit domain list |
View and customize allowed/blocked domains |
AntiVibe Safety Guards — Open settings |
Jump to the VS Code settings section |
Token Optimization: Cut Token Usage 30-60%
Claude Code sessions burn tokens on redundant file reads, verbose output, and blind navigation. Token Optimization is a single toggle in the Intelligence tab that activates six coordinated cost-reduction techniques:
| Technique |
How it saves tokens |
| Project dependency map |
Scans your workspace, parses imports/exports, and generates .claude/project-map.md — a structural graph so Claude navigates without trial-and-error reads |
| Verbosity control |
Merges concise-mode instructions into CLAUDE.md — 4 levels from normal to json-only |
| Read deduplication |
Hook tracks file reads per session; warns when re-reading unchanged files within a configurable window (default 5 min) |
| .claudeignore |
Gitignore-style patterns that advise Claude to skip irrelevant files (node_modules, lock files, build output, minified bundles) |
| Output compression |
Hook compresses verbose CLI output (git, npm, test runners) — deduplicates repeated lines, extracts summaries, caps at configurable max lines |
| Context budget watchdog |
Tiered alerts (70%, 85%, 95%) when context window fills up — with actionable prompts to run /compact |
One click to enable. After toggling on, restart VS Code for hooks to activate. Disable reverses everything cleanly — removes hooks, CLAUDE.md block, and settings.
How it works
- CLAUDE.md merge block — idempotent HTML comment markers inject concise-mode instructions; no manual editing needed
- Python hooks — lightweight PreToolUse and PostToolUse hooks in
~/.claude/settings.json that never block Claude (advisory only, always exit 0)
- CLAUDE.md Analyzer — detects oversized sections, duplicate content, and estimates per-section token cost with actionable recommendations
- Secure by default — temp files use restricted permissions (0o700/0o600), session IDs are sanitized against path traversal, pattern lengths are bounded to prevent ReDoS
Settings
14 granular settings under claude-code-toolbox.tokenOptimization.* — enable/disable each technique independently, configure thresholds, file extensions for the project map, compression limits, and more.
Commands
| Command |
What it does |
Token Optimization — Enable |
Activates everything: merges CLAUDE.md block, generates project map, installs hooks |
Token Optimization — Disable |
Cleanly removes all hooks, CLAUDE.md block, and artifacts |
Token Optimization — Generate project map |
Re-scans workspace and writes .claude/project-map.md |
Token Optimization — Analyze CLAUDE.md |
Opens a token analysis report with per-section breakdown and recommendations |
Token Optimization — Create .claudeignore |
Generates a .claudeignore with sensible defaults |
Token Optimization — Status |
Shows current configuration and hook installation state |
Intelligence: Context & Migration Hub
One Click Setup migrates your team from Cursor and/or Copilot to Claude Code automatically — ports MCP servers, syncs rules to CLAUDE.md, scaffolds memory bank, and migrates skills. All using bundled CLIs (no network npx fetch).
Thinking Machine Mode primes Claude Code with MCP & skills awareness, context packs, and readiness checks — making "refresh what this repo knows" a single action.
MCP & Skills Hub
Browse and manage MCP servers and skills from one place. Reads Claude Code's native config (~/.claude.json for user, .mcp.json for project). Registry browse, install, stash/hide — no more raw JSON editing.
Workspace Kit
A visual checklist for everything Claude Code needs: rules, memory bank, CLAUDE.md, mcp.json — see what exists and what's missing at a glance.
Getting Started
- Install from the VS Code Marketplace
- Click Claude Code ToolBox in the Activity Bar (left icon column)
- Open MCP & skills from the Side Bar
7 Preset Swarm Commands
Ship with the starter pack — each dispatches all team agents in parallel:
| Command |
What it does |
/plan-team |
Architect + PM plan, you approve, devs build |
/debate-team |
Agents debate approaches, judge picks the winner |
/review-team |
Multi-perspective code review |
/security-team |
Threat modeling + vulnerability scan |
/sdlc-plan-then-code |
Full SDLC: plan → approve → implement → test → review |
/refactor-team |
Parallel refactoring with cross-review |
/spec-team |
Spec generation with multi-agent validation |
Requirements
| Requirement |
Notes |
| VS Code |
1.99+ |
| Claude Code |
Install the Claude Code extension |
| Node.js |
20+ for bundled CLIs |
Key Settings
| Setting |
Purpose |
claude-code-toolbox.agentTeams.* |
Model, protocol, max concurrent agents, cost cap |
claude-code-toolbox.agentDashboard.* |
Enable/disable dashboard, hook port, safety alerts |
claude-code-toolbox.tokenOptimization.* |
Verbosity level, project map, read dedup, output compression, .claudeignore, context budget |
claude-code-toolbox.oneClickSetup.* |
Migration tracks (Cursor, Copilot) |
claude-code-toolbox.intelligence.* |
Context pack, auto-scan, notepad |
claude-code-toolbox.thinkingMachineMode.* |
Priming and awareness behavior |
Works on macOS, Windows, and Linux. Full JetBrains parity ships as a separate IntelliJ plugin with the complete Agentic Teams runtime.
Disclaimer
Independence and trademarks. Claude Code ToolBox is independent community tooling. It is not affiliated with, endorsed by, or maintained by Microsoft, GitHub, Cursor, OpenAI, Anthropic, or other vendors. Product names may be trademarks of their respective owners.
MIT "AS IS". Licensed under the MIT License. Not professional services, security audit, or legal review.
Your responsibility. Backups, secrets hygiene, and policy compliance are yours.
License
MIT — Copyright (c) 2026 amitchorasiya.
