ClaudePacer Documentation

Installation

ClaudePacer is a Chrome extension — no account, no API key, no setup wizard required.

1. Visit the Chrome Web Store and search "ClaudePacer" (or use the direct link from the download page).
2. Click Add to Chrome → Add extension.
3. Open claude.ai. The side panel opens automatically.

First Run

When you first open claude.ai after installing, ClaudePacer starts in Demo Mode — showing realistic fake data (420 tok/min, 38% used) so you can see the product immediately.

Once you send your first message, the extension detects a real API response and switches to Live mode automatically. The green dot lights up.

Demo Mode

Demo mode activates when no real token data has been received yet in the current session. It shows:

• 420 tok/min burn rate
• 38% weekly budget remaining
• A "Demo" label in the header

Demo mode is intentional — it makes the extension look great immediately on install, so you can explore the UI before sending any messages.

Speedometer — Burn Rate

The speedometer shows your current token burn rate in tokens per minute. It's calculated from a rolling 60-second window of token activity.

Color zones

• Green (0–800 tok/min) — relaxed pace, you have plenty of runway
• Amber (800–2000 tok/min) — active session, monitor your gauge
• Red (2000+ tok/min) — burning fast; large inputs or long responses

Why does the needle sometimes jump?

Large input messages (e.g. pasting a whole file into a chat) register as a spike. The needle smooths out as the rolling window advances.

Budget Gauge — Weekly / Session

The arc gauge on the right shows how much of your budget remains. Toggle between Weekly and Session view using the buttons above the gauge.

Weekly view

Shows % remaining of your total weekly token budget. Resets on the day configured in Settings (default: Monday). When Anthropic sync is active, this reflects actual usage data from Anthropic's servers.

Session view

Shows % remaining of the current 5-hour session window. Claude Pro/Max plans reset every ~5 hours. This view helps you pace within a single session.

Limit Prediction

Below the gauges, ClaudePacer shows a forward-looking prediction:

This is calculated by dividing your remaining budget by your current rolling burn rate. It updates every 30 seconds. Idle periods cause it to extend; heavy sessions cause it to shrink.

Anthropic Sync

ClaudePacer periodically calls Anthropic's usage endpoint to get your actual plan usage. When this succeeds, the banner shows:

Click re-sync to force a refresh. If sync fails (e.g. you're logged out), the gauge falls back to estimated values.

Compress This Chat is a one-click token saver. It sends Claude a structured compression request that:

1. Summarizes your entire conversation history
2. Preserves all key decisions, code snippets, and open threads
3. Saves the result to your MEMORY.md file
4. Saves 30–55% of context tokens on average

When to use it

Use Compress This Chat when:

• Your conversation has gone on for 20+ turns
• You notice the speedometer climbing (long context = more input tokens per turn)
• You're approaching your session limit and want to continue in a fresh chat

What it sends to Claude

ClaudePacer injects this compression prompt into your active chat:

Please compress this conversation into a dense context block for MEMORY.md.
Include: all key decisions made, code changes agreed on, open questions,
and the current task state. Format as markdown. Be ruthlessly concise.

Open Fresh Chat does two things in sequence:

1. Saves — runs the compression prompt and saves the output to MEMORY.md
2. Resets — opens a new claude.ai conversation with zero token overhead

In the new chat, you paste your MEMORY.md summary and Claude picks up exactly where you left off — with a completely fresh context window.

Token savings

A long coding session might accumulate 40,000+ input tokens of context. Starting fresh with a 2,000-token summary means each subsequent message costs 95% less in input tokens.

MEMORY.md is a markdown file in your project root (or ~/.claude/) that stores compressed conversation context. It's the same convention used by Claude Code's built-in memory system.

File structure

# Project Memory

## Active Task
Building a React dashboard with authentication...

## Key Decisions
- Using Supabase for auth (not NextAuth)
- API routes in /app/api/
- Deployed to Vercel

## Open Questions
- Should we add rate limiting to the API?

## Last Updated
2026-04-14 by ClaudePacer

Auto-Update Memory (Pro)

When enabled in Settings, ClaudePacer automatically runs a lightweight compression after each session and updates MEMORY.md. You never have to think about it — context is always preserved.

Desktop Bridge — Setup

The Claude Code bridge lets the extension track token usage from the Claude Code terminal CLI, not just claude.ai. If you only use claude.ai in the browser, you don't need this.

Requirements

• Node.js installed (node --version should work)
• Claude Code CLI installed (claude --version)
• macOS or Linux (Windows support planned)

Start the bridge

Option A — Terminal:

cd ~/Documents/ClaudePacer/claude-bridge
/usr/local/bin/node bridge.js

Option B — Double-click: Find start-bridge.command in Finder and double-click it.

Desktop Bridge — Usage

When the bridge is running and the extension detects it, a chip appears in the side panel:

Token usage from your terminal sessions is merged into the same speedometer and budget gauge. The weekly total reflects all Claude usage — browser + desktop.

Stopping the bridge

Press Ctrl+C in the terminal where the bridge is running. The extension will show the chip as disconnected after the next polling interval.

Token Accuracy

ClaudePacer's token counts are estimated, not exact. Here's why and how accurate they are in practice:

The technical challenge

Claude.ai's internal SSE stream strips the usage fields that Anthropic's standard API includes. There's no input_tokens or output_tokens field in the stream — only raw text deltas.

How estimation works

• Output tokens: Accumulated text from content_block_delta events ÷ 4
• Input tokens: POST request body text length ÷ 4
• Gas gauge: Synced from Anthropic's actual usage API (exact) when available

Practical accuracy

The ÷4 heuristic is widely used and generally within 10–15% of actual token counts for English text. For code or non-English text, variance may be higher. The speedometer trend is reliable even when absolute numbers aren't perfect.

Settings

Open Settings from the gear icon in the side panel, or by right-clicking the extension icon → Options.

Available settings

• Weekly token limit — set your plan's total weekly budget
• Alert threshold — get warned when you've used X% of your budget
• Week start day — choose which day your week resets (default: Monday)
• Session usage sync — enable/disable Anthropic API sync for the gas gauge

Troubleshooting

Side panel doesn't open

Click the ClaudePacer icon in Chrome's toolbar. If you don't see it, click the puzzle piece icon and pin ClaudePacer.

Gauges stuck in Demo mode

This means no real API responses have been detected. Try: reload the claude.ai tab, send a message, and wait 5 seconds. If still stuck, go to chrome://extensions → click ↺ on ClaudePacer to reload the extension.

Green dot stays grey (Idle)

The dot only goes green during an active streaming response. If you're between messages, idle is correct. It will update within seconds of your next message.

Bridge not connecting

Make sure the bridge is running (node bridge.js in the claude-bridge/ folder). Check that nothing else is using port 7823. The extension polls every 60 seconds, so wait up to a minute after starting the bridge.

Token counts seem very high

In project-heavy chats, the entire project context is sent with each message. A project with 50 files can add 100,000+ input tokens per turn. This is real — ClaudePacer is accurately showing you the cost of large project contexts.

Privacy

ClaudePacer is designed to be private by default:

• Your conversation content is never read or stored — only text length is measured
• Token counts are stored in chrome.storage.local — only in your browser
• No analytics, no crash reporting, no telemetry
• The bridge only reads .jsonl file sizes and timestamps — not content
• The Anthropic sync call is made from your browser using your existing session — no credentials are stored by ClaudePacer

See the full Privacy Policy for details.