Token check and savings

The token check is the free first step in AgentWitch. It scans your computer for the rules, skill lists and tool servers your coding assistants send with every model call, before you type a word, and it shows what each one costs. You can then trim the biggest ones in a few clicks and watch your savings in the project.

#What the check looks at

A coding assistant starts every model call with some fixed text. The check measures three kinds of it for each assistant it finds (Cursor, Claude Code, Codex and Antigravity):

Always-on rules. Rule and instruction files that load on every call, such as Cursor rules set to always apply, CLAUDE.md, AGENTS.md and GEMINI.md. It reads them in your project folder and in your home folder.
The skill list. The assistant sees one short line for every skill it has, so a long list costs tokens even if you never use most of them.
MCP servers. Each enabled MCP server sends its full tool list with every call, even when no tool is used.

The numbers are estimates, marked with ≈. AgentWitch counts the size of the text and divides by four. It is not your provider bill. After the first call of a run, most of this text is read from the cache at a lower price, so a run costs less than the number of calls times the total.

Some things are not counted. Cursor user rules typed into Cursor Settings are not read. MCP tool lists are measured only when Cursor has cached them; other servers show "size unmeasured (not counted)".

AgentWitch also adds a small amount of its own text, so assistants can find skills and save what they learn. The check shows it on its own line ("AgentWitch adds ≈… per model call, on purpose") and never counts it as a leak.

#Run the check

On the AgentWitch website, choose Run a free token check. Sign in and follow the setup: connect your computer, pick your AI tool, then choose a project folder.
On Home, find your project in the table. The Setup tokens per run column says Not scanned yet until the first scan. Choose Scan token leaks in that row.
Wait for Scanning… to finish. The row then shows a number, or Lean when almost nothing can be trimmed.
Choose that number to open Health › Token leaks for the project.

On a phone, choose Send the install link to your Mac and open the link on your computer. The scan runs on your computer, so open the project in a browser on that computer. If the scan says there is no folder for the project, open Health › Folder, add the path to your project and scan again.

The scan only runs when you start it, never on a schedule. Scan now is hidden for 15 minutes after a scan.

#Read the result

The top card shows one big number: the tokens that load into every model call and could go. Under it you see what share of the total that is. Next to it, bars show where the weight comes from: Always-on rules, Skill list and MCP tools, split into Needed and Removable, up to.

If you use more than one assistant, you get one tab per assistant with its own total and the number of places to trim. All adds them up. "Up to" is deliberate: it is the most you could save, not a promise. If nothing is worth trimming, the card says "Nothing to trim. Your setup is lean."

Below the card, places to trim, biggest first lists each leak. A file is flagged when it is about 8,000 tokens or more. The skill list becomes a leak at about 1,000 tokens. Open one to see the problem and the fix.

#Fix the leaks

Choose Fix N leaks one by one. Each step is titled "Leak 1 of N" and offers a fix:

Heavy rule files. AgentWitch proposes a change, such as splitting or shortening the file. Choose Apply and the file changes on your computer. The step then says "Fixed. Each run is now ≈… lighter." If there is no automatic fix, it points you to the files so you can edit them, or choose Open in Skills.
MCP servers. The step lists the enabled servers. Remove the ones you do not use in Cursor Settings › MCP, restart Cursor, then choose I've done this.
Skill list. Remove skill folders you never use, then choose I've done this. This deletes the files from your computer.

Use Back, Skip and Next leak to move around. See result shows Before and Each run now, the percentage saved and the status of each leak: Fixed, Skipped or Not done.

MCP and skill-list steps are offered for Cursor only. For Claude Code, Codex and Antigravity, the check lists rule files. Viewers see the numbers but not the fix buttons. The estimate updates as soon as you apply a rule fix, and a fresh measurement comes with the next scan.

Open Everything that loads at the bottom to see all rules, MCP servers, skill folders and assistant storage. You can clean old chat history there after you confirm. That cannot be undone.

When the check finds leaks, you can also copy a ready-made post about your result (Copy a post for X, LinkedIn, Reddit or Slack) or choose Invite a teammate with this result. Nothing is posted for you.

#Heavy runs and savings

Heavy runs. In History, a run that used a lot of tokens shows Heavy. Open it for a Token health check: the likely causes, such as a very large setup on every step, many steps, the same file read again and again, a vague request or an unused plugin. You can change the line above which a run counts as heavy at the top of History. The default is 400k tokens. A very heavy turn can also get a one-line note in your assistant's chat.

Savings. Savings show in three places:

Overview has What AgentWitch did for you: tasks done, runs with project notes, mistakes avoided and tokens saved (est.) for the last 7 days.
Knowledge impact shows how many runs reused earlier work, which skills were used most, and how many notes each computer learned over 7, 30 or 90 days.
Skills › In use shows, for each skill, how often it ran and the estimated tokens it saved.

A skill's saving compares its runs with a baseline. The first run sets the baseline, and about one use in ten skips the skill on purpose to keep the baseline fair. Until a skill has three samples, its number carries ≈. In general, ≈ marks an estimate and everything else is measured.

For how skills get made, see Skills and scripts.

#If something does not work

"Scan runs on this computer." Open the project in a browser on the computer where AgentWitch Local runs.
"AgentWitch Local on this computer is out of date." Choose Update this computer.
"No scan yet." Add the project path under Health › Folder, then choose Scan now.
MCP sizes missing. Open the server once in Cursor so it caches its tools, then scan again.

#Query aliases

run a free token check
what loads on every call
reduce tokens used by rules and MCP servers
fix token leaks
how are tokens saved measured