nudge-mcp

<!-- mcp-name: io.github.jwill824/nudge-mcp -->

MCP server that exposes Claude Code and GitHub Copilot CLI session token usage and cost data as tools, so your AI assistant can query its own usage mid-conversation.

Prerequisites

• Claude Code and/or GitHub Copilot CLI installed
• uv installed

Installation

1. Register the MCP server

#### Claude Code

claude mcp add nudge-mcp -- uvx nudge-mcp

Or add manually to ~/.claude/settings.json under mcpServers:

"mcpServers": {
  "nudge-mcp": {
    "type": "stdio",
    "command": "uvx",
    "args": ["nudge-mcp"]
  }
}

#### GitHub Copilot CLI

Create (or update) .mcp.json in the repo root:

Published package (default):

{
  "mcpServers": {
    "nudge-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["nudge-mcp"]
    }
  }
}

Local development:

{
  "mcpServers": {
    "nudge-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "/path/to/nudge-mcp", "nudge-mcp"]
    }
  }
}

Note: .mcp.json is gitignored because it may contain machine-specific config. Each contributor creates their own locally.

2. Restart Claude Code

The MCP server connects on startup. You should see nudge-mcp listed when you run /mcp in Claude Code.

MCP Tools

Once installed, ask Claude naturally — it will call the tools automatically.

claude_session_report

Recent sessions with per-session cost and efficiency metrics.

"Show my session report for today"
"Show the last 10 sessions"
"Show sessions for April 2026"

claude_monthly_summary

Total token usage and estimated cost for a Claude Code billing month. Shows spend vs your configured monthly budget with remaining runway.

"What's my estimated spend for this month?"
"Show my April 2026 monthly summary"

copilot_monthly_summary

Monthly output token summary for GitHub Copilot CLI sessions. Shows total output tokens, top projects, and equivalent API cost vs your flat subscription to highlight savings.

"What's my Copilot usage this month?"
"Show Copilot summary for April 2026"

configure_subscription

Update the active Claude Code or GitHub Copilot plan and monthly budget.

"Set my Claude plan to claude_max_400"
"My Copilot budget changed to $39 — update it"
"Set my Claude budget to 350"

Claude plans: claude_pro ($20), claude_max_100 ($100), claude_max_200 ($200), claude_max_400 ($400), api ($0)

Copilot plans: copilot_free ($0), copilot_pro ($10), copilot_pro_plus ($39), copilot_business ($19/seat), copilot_enterprise ($39/seat)

claude_tool_impact

Compare Claude Code session efficiency between sessions that used a specific tool vs those that didn't.

"What impact does Serena have on my session efficiency?"
"Compare sessions that used ck vs those that didn't"
"Show the impact of Read calls in April 2026"

copilot_session_report

Show GitHub Copilot CLI session output token usage and efficiency metrics.

"Show my Copilot CLI session report"
"Show Copilot sessions for April 2026"
"Show today's Copilot sessions"

Note: Only output_tokens are tracked. The Copilot CLI does not expose input or cache token counts.

claude_calibrate_pricing

Update the discount factor from your actual Claude Code billing statement.

"Calibrate pricing — I was billed $185.50 last month"

copilot_behavior_report

Cross-session pattern analysis. Aggregates behavioural signals across recent sessions to surface recurring inefficiencies with actionable recommendations. Includes aggregate context cost (heavy tool results), store_memory utilisation, and a smart tool recommendation when heavy view usage is detected without Serena, ck, or ast-grep.

"Analyse my last 10 Copilot sessions for bad habits"
"Show behaviour report for April 2026"

analyze_copilot_session

Deep-dive analysis of a single Copilot CLI session. Checks prompt quality, tool batching (parallelism), bash overuse vs specialised tools, memory utilisation, context cost (heavy tool results), smart code intelligence tool usage (Serena, ck, ast-grep), MCP tool budget (avg KB/call vs call frequency with disable recommendations), and agentic workflow tools — structured systems like Superpowers, Get Shit Done (GSD), and Spec-Kit that bring tools, subagents, and multi-step workflows to reduce context overhead and prevent context rot. When no workflow tools are detected, the analysis gives personalized recommendations based on session characteristics.

"Analyse my current session"
"Analyse session abc123 for inefficiencies"

copilot_premium_usage

Fetch live Copilot premium request usage from the GitHub API. Shows requests used and cost broken down by model, with progress against your configured overage budget. Uses the gh CLI token automatically — no extra setup needed if you're already authenticated with gh.

Requires a fine-grained PAT (or gh auth login) with Plan (read) permission. Only returns data for individual Copilot plans, not org/enterprise-managed licenses.

"Show my Copilot premium request usage for this month"
"How many premium requests have I used in April 2026?"

copilot_tool_impact

Analyze how a specific tool affects Copilot CLI session efficiency — comparing output tokens/turn and estimated cost between sessions that used the tool and those that didn't. The Copilot equivalent of tool_impact. Useful for measuring the real-world impact of tools like Serena, ck, ast-grep, or any MCP tool once you start using them.

"Show the impact of serena on my Copilot sessions"
"How much does ck reduce my output tokens per turn?"
"Compare bash-heavy vs non-bash-heavy sessions"

record_copilot_spend

Manually record your actual Copilot overage spend for a month (from the GitHub billing UI). Used when the live API is inaccessible (org/enterprise-managed plans). The recorded amount is displayed in copilot_monthly_summary and used for budget forecasting.

"Record my Copilot spend of $18.00 for April 2026"

copilot_budget_forecast

Forecast end-of-month Copilot overage spend using your current burn rate, and estimate how much you could save by fixing behavioural inefficiencies (low batching, bash overuse, vague prompts).

Requires record_copilot_spend and configure_subscription (with overage_budget) to be set first.

"Forecast my Copilot budget for this month"
"How much will I spend by end of April at this rate?"
"How much would I save if I fixed my batching habits?"

copilot_model_efficiency

Analyse how well the active model matched task complexity across Copilot CLI sessions. Scores each turn 0–8 across four signals (prompt length, output tokens, tool call count, unique tools), classifies each turn as over-powered / fit / under-powered, and estimates savings from switching simple turns to a budget model.

"Has my model usage been efficient this month?"
"Show me which sessions were over-powered"
"Analyse model efficiency for the last 10 sessions"

Pricing Model

Costs are calculated per-model using Anthropic API list prices, then scaled by a discount_factor:

The model used is detected per-turn from the JSONL (message.model), so Opus sessions are priced at Opus rates automatically.

The discount_factor (default 0.5868 ≈ 41% off list) approximates subscription value vs API list price. Recalibrate monthly with claude_calibrate_pricing for accuracy.

Testing with MCP Inspector

npx @modelcontextprotocol/inspector uvx nudge-mcp

Interpreting the Stats

Cache hit %

tokens/turn

Data

Claude Code session data is read directly from ~/.claude/projects/. No setup required — sessions are parsed from JSONL files that Claude Code writes automatically. The model used is detected per-turn from message.model in the JSONL, enabling accurate per-model cost calculations.

Copilot CLI session data is read from ~/.copilot/session-state/.

See CONTRIBUTING.md for development workflow and project structure.