Claude Token Dashboard

An MCP server for Claude Code that shows every MCP server, skill, command, agent and plugin on your machine, estimates how many tokens each one adds to every request, and lets you switch them off per project or for all projects. A live tracker reads your Claude Code sessions as they happen and counts the tokens your switched-off items no longer send.

No dependencies. Needs Node.js 20 or newer.

Add it to Claude Code

Every project (recommended)

Add it once at user scope and it loads in every project you open:

# macOS / Linux
claude mcp add --scope user token-dashboard -- npx -y github:Yiamin24/claude-token-dashboard

# Windows
claude mcp add --scope user token-dashboard -- cmd /c npx -y github:Yiamin24/claude-token-dashboard

From a local copy

git clone https://github.com/Yiamin24/claude-token-dashboard.git
claude mcp add --scope user token-dashboard -- node /full/path/to/claude-token-dashboard/server.js

One project only

Run this inside the project, or leave out --scope user from either command above:

claude mcp add token-dashboard -- npx -y github:Yiamin24/claude-token-dashboard

When a Claude Code session starts, the server also starts the dashboard at http://127.0.0.1:7377/ for the project you're in. Open it yourself, or ask Claude to "open the token dashboard".

To remove it: claude mcp remove --scope user token-dashboard

To run only the dashboard, without Claude Code:

npm start        # node server.js --standalone

Works across your AI tools

It also reads the MCP servers and skills of other AI coding tools on the same machine. That covers both global and project-level config, so you can see everything that loads, wherever it's set up. When the same server or skill is set up in several tools, it shows as one card with a switch per tool.

"Read-only" means the tool keeps its on/off state inside the app instead of in a file, so the dashboard shows the item and its cost but you switch it in that tool. Skills of other tools are read-only too, because those tools have no per-skill setting. Every file the dashboard writes is backed up first, and only the one setting is changed.

Projects: the project list comes from every tool, not just Claude Code: Claude Code's project list and session history, Codex's trusted projects, Antigravity's projects, and the workspace history of VS Code, Kiro, Cursor and Windsurf. It also includes folders under Desktop and Documents that have their own AI-tool setup. Projects with their own MCP servers or skills are listed on the overview; click one to see what it adds.

Measuring a server once covers every tool that runs the same server. Antigravity caches each connected server's tool list on disk, and the dashboard reads that cache, so those servers are measured without being started.

From Claude: context_report takes agent (codex, gemini, … or all). context_toggle takes agent too, and agent: "all" switches the same server in every tool that allows it.

Auto mode

Auto mode switches Claude Code's MCP servers and skills to fit each project, and optionally the task you're on.

• By project: it reads each project's dependencies and files. For example, next + tailwindcss means a web front end, build.gradle means Android, .sql/.ipynb files mean data work, binaries mean reverse engineering, and .docx/.pptx files mean Office documents. It then classifies every server and skill by its name, description and tools, and switches off the ones that project doesn't need.
• By task: tell it what you're working on, either on the Auto mode page or by asking Claude (context_auto with task). For example, "reverse engineer this apk" brings rea back for that project.
• Ahead of time: Claude Code reads these settings when a session starts. With auto mode on, every known project is prepared in the background (on start, every 10 minutes, and when something new is installed), so the right set is already there when you open a project.

How careful it is:

• Off by default. It shows a preview of every change and the reason for it.
• Gentle (default) keeps unneeded skills listed by name only, so Claude can still use them if asked. Strict switches them off completely. Unneeded MCP servers are switched off in both modes.
• It leaves things alone when it isn't sure: items it can't classify, process skills useful anywhere (planning, debugging, reviews, git), and anything you switched yourself, including later changes to something auto mode set.
• It only brings back what it switched off itself. Undo puts back everything it changed in a project.
• It writes Claude Code's per-project .claude/settings.local.json only. For other tools it lists the servers that aren't needed so you can switch them there.
• You can exclude a project with "Never run auto mode in this project".

Learning: gets better every day

Auto mode learns from how you actually work, so its decisions improve with every session.

• What you use: it reads your Claude Code sessions and records which skills, MCP servers and slash commands get used, per project. On first start it reads your last 7 days, so it's useful right away. Anything you've used in a project in the last 30 days stays on there.
• Your fixes: if you switch back on something auto mode turned off, that counts as a correction. The item stays on in that project, and auto mode leaves it alone from then on.
• Kinds of work: an item used mostly in, say, data projects is learned as a data tool, even if its description doesn't say so. One used across very different projects is learned as useful anywhere.
• Never used: once it has seen 20+ sessions over at least a week, skills that no project has used drop to name-only (Gentle) or off (Strict). Unused MCP servers are switched off only in Strict mode.

The Learning page shows what it learned: what you use here and everywhere, and what loads on every request but is never used, with the token cost.

AI helper (optional)

Some servers, skills, tasks and projects don't match any keyword rule. The AI helper sorts those into the same categories. It's off by default, every answer is saved (each item is asked about once), and the Learning page shows every answer and the total tokens used.

• Local (Ollama): free, and nothing leaves your computer. Install Ollama and pull a model (ollama pull llama3.2).
• Claude API: uses the official @anthropic-ai/sdk (installed as an optional dependency by npm install). Credentials come from ANTHROPIC_API_KEY or an ant auth login profile, and the dashboard never stores them. The default model is claude-opus-5-5, and you can pick claude-sonnet-5-5 or claude-haiku-5-5 on the Learning page. Requests use structured JSON output, low effort, and server-side refusal fallback (fallbacks: "default").

Learning data is stored in ~/.claude/token-dashboard/learned.json on your computer.

What it does

MCP schema counting

Claude Code loads MCP tool definitions in one of two ways:

• full: every tool's full schema is sent with each request.
• deferred: only tool names and server instructions are sent; schemas load when a tool is needed.

Claude Code switches to deferred when MCP tools would fill more than about 10% of the context window. auto follows the same rule; you can force either mode from the MCP tab.

MCP tools (kept small: about 500 tokens)

• dashboard: opens the dashboard and returns its URL.
• context_report: a text summary of what's loaded, the costs, savings so far and new items.
• context_toggle: { name, state: on|off|name-only, scope: project|global, agent }.
• context_auto: { task, apply, undo }. Previews (or applies) auto mode for this project.

Notes

• Changes apply to new Claude Code sessions, not the one already running.
• Token counts are estimates at about 4 characters per token.
• Servers that need OAuth sign-in (for example remote HTTP servers) can't be measured from outside Claude Code. They show "needs sign-in to measure".
• claude.ai connectors added from your claude.ai account aren't stored locally, so they aren't listed.
• Plugin SessionStart hooks can add text to each session. They're flagged on the plugin but not counted.
• Only one dashboard runs per machine. Extra Claude sessions share it, and if the session hosting it closes, another one takes over within about 20 seconds.
• Settings: TOKEN_DASHBOARD_PORT (default 7377), TOKEN_DASHBOARD_DATA (default ~/.claude/token-dashboard), CLAUDE_CONFIG_DIR.

Files

server.js        MCP stdio server + --standalone entry
src/scan.js      finds items, works out on/off state and token cost, spots new items
src/agents.js    other AI tools: Codex, Gemini, Antigravity, Kiro, Cursor, VS Code...
src/auto.js      auto mode: project/task detection, plan, apply, undo
src/learn.js     learning: usage per project, corrections, AI answers
src/ai.js        optional AI helper (Ollama or Claude API)
src/toggle.js    writes switches to Claude Code settings (with backups)
src/probe.js     measures MCP servers (stdio and HTTP)
src/tracker.js   live transcript tailing and the savings ledger
src/web.js       local HTTP server, live event stream, file watchers
public/index.html  the dashboard
test/            node --test suites (npm test)

License

MIT © 2026 Yiamin24. See LICENSE.