Tool Curation

July 6, 2026 · View on GitHub

Overview

The Tool Curation feature analyzes the gap between the tools available in your environment and the tools actually used within a configurable look-back window. It surfaces actionable recommendations to reduce prompt overhead caused by idle MCP servers and stale skill files.

The curation section appears in the Tools tab of the Usage Analysis Dashboard.

Why It Matters

Every MCP server and skill file contributes a description block to every prompt — even when the tool is never called. Over time, accumulating unused tools inflates prompt token counts, increases latency, and can hit context-window limits.

Tool Curation gives you a concrete view of which tools are pulling their weight and which ones can be safely disabled or refined.

What Is Analyzed

Available Tools (discovered on disk)

MCP Servers

The extension reads the following config files to build the list of configured MCP servers (checked in priority order, deduplicated by server name):

Location	Scope	Editor
`<workspace>/.vscode/mcp.json`	Workspace	VS Code
`<workspace>/.mcp.json`	Workspace / repo root	Visual Studio
`<workspace>/.vs/mcp.json`	Solution-scoped	Visual Studio
`<workspace>/.cursor/mcp.json`	Workspace	Cursor
`%USERPROFILE%/.mcp.json` (Windows) / `~/.mcp.json` (macOS/Linux)	User-global	Visual Studio

All files share the same JSON schema: { "servers": { "<name>": { ... } } }.

Skill Files

Skill files (SKILL.md) are discovered from the following directories:

Workspace-scoped (relative to each workspace folder):

Directory	Editor
`.github/skills/`	VS Code / GitHub Copilot
`.claude/skills/`	Visual Studio
`.agents/skills/`	Visual Studio

User-scoped (in the home directory):

Directory	Editor
`~/.copilot/skills/`	Visual Studio
`~/.claude/skills/`	Visual Studio
`~/.agents/skills/`	Visual Studio

Each skill entry's description is extracted from the first description: line or # Heading in SKILL.md.

Extension/Runtime Tools

Tools registered via vscode.lm.tools at runtime are enumerated separately. MCP tools discovered this way are matched back to their server name via the mcp__<server>__<tool> naming convention.

Used Tools (from session analysis)

Tool usage is counted from the look-back window configured via the aiEngineeringFluency.curation.timeWindowDays setting (default: 30 days, options: 7 / 30 / 90).

MCP server usage: counted from mcpTools.byServer in the aggregated usage period
Tool call usage: counted from toolCalls.byTool across all sessions in the window

Setting	Type	Default	Description
`aiEngineeringFluency.curation.timeWindowDays`	`number`	`30`	Look-back window for determining "used" vs "unused" tools. Options: `7`, `30`, `90`.

Commands

Command	Description
`aiEngineeringFluency.openMcpJson`	Opens the first MCP config file found in the workspace (checks `.vscode/mcp.json`, `.mcp.json`, `.vs/mcp.json`, `.cursor/mcp.json` and user `~/.mcp.json` in priority order; offers to create `.vscode/mcp.json` if none exist).

Architecture

The curation logic lives in src/toolCuration.ts. The module is intentionally pure (no direct VS Code API calls) so it can be reused by the CLI.

Function	Purpose
`enumerateRuntimeTools(tools)`	Converts `vscode.lm.tools` into `AvailableToolEntry[]`
`parseMcpJson(path)`	Parses a single `mcp.json` file and returns server names
`buildMcpEntriesFromJson(folderPaths)`	Discovers all MCP entries from all config locations
`discoverSkillEntries(folderPaths)`	Discovers all skill entries from workspace + home dirs
`analyzeToolCuration(available, usage, windowDays)`	Runs the full curation analysis

Results are surfaced through:

The Tools tab in the Usage Analysis webview (src/webview/usage/main.ts)
The Insights engine (src/insightsEngine.ts) — adds unused-mcp-servers and stale-skills insight cards
The CLI curation command (cli/src/commands/curation.ts)

CLI Usage

copilot-token-tracker curation
copilot-token-tracker curation --json          # machine-readable output
copilot-token-tracker curation --days 7        # use 7-day window