Claude Token Manager

A VS Code extension that adds two missing features to the Claude Code VS Code extension:

1. Token Usage Display — a numeric status bar item showing exactly how much of Claude's context window is in use, with a detailed breakdown panel.
2. Automatic .md Memory System — watches your Claude sessions passively, auto-summarizes them using the Anthropic API, and builds a living memory file that you can paste at the start of any new session so Claude never re-learns the same things twice.

Requirements

• VS Code 1.85.0 or later
• The Claude Code VS Code extension installed and in use
• Node.js 18+ (to build from source)
• An Anthropic API key (only required for the auto-summarization feature)

Installation (from source)

# 1. Clone / download this folder
cd claude-token-manager

# 2. Install dev dependencies
npm install

# 3. Compile TypeScript
npm run compile

# 4. Package as a .vsix
npm run package

# 5. Install in VS Code
code --install-extension claude-token-manager-0.1.0.vsix

Or press F5 inside VS Code with this folder open to launch an Extension Development Host for testing.

Setup

API Key (required for auto-summarization)

On first activation you will see a notification prompting you to add your Anthropic API key.
Click Add API Key and paste your key (starts with sk-ant-).

The key is stored in VS Code's encrypted SecretStorage — never in settings.json or any plain-text file.

You can update it at any time via the Command Palette:

Claude Token Manager: Set API Key

Feature 1 — Token Usage Display

How it works

The extension watches ~/.claude/projects/{your-workspace}/ for .jsonl files that Claude Code writes during sessions. It parses those files passively (read-only, never writes) and aggregates token counts.

Status bar

A persistent item in the right side of the status bar shows:

⚡ 42,310 / 200,000 tokens

Updates within 2 seconds of any Claude Code activity. Shows ⚡ No active Claude session when idle.

Token usage panel

Click the status bar item (or run Claude Token Manager: Show Token Usage Panel) to open a webview showing:

• Tokens used vs. remaining (progress bar)
• Input / output / cache-read token breakdown
• Rough cost estimate at Claude Sonnet 4.5 pricing ($3/M input, $15/M output)
• Bar chart of token usage across the last 20 conversation turns
• Clear Session View button

Note: The panel shows data from the most recently active .jsonl file. The currentContextSize shown in the status bar is the input_tokens value from the most recent assistant turn — this is the actual context window fill reported by the API, not a running total.

Feature 2 — Automatic Memory System

File structure

On first activation the extension creates a claude-memory/ directory in your workspace root:

your-project/
└── claude-memory/
    ├── MEMORY.md      ← auto-updated, copy-paste this into new sessions
    ├── BUG_FIXES.md   ← append-only log of every bug resolved
    ├── DECISIONS.md   ← append-only log of architectural decisions
    └── SUMMARY.md     ← human-facing overview (not injected)

Auto-summarization

After a Claude Code session goes idle for 60 seconds (no new writes to the .jsonl file), the extension:

1. Reads the session transcript from the .jsonl file
2. Sends it to claude-haiku-4-5-20251001 (the cheapest model — summarization doesn't need Sonnet)
3. Extracts: project context, tech stack, conventions, key files, bug fixes, decisions, open issues
4. Appends new bug fixes to BUG_FIXES.md and decisions to DECISIONS.md
5. Rewrites MEMORY.md to stay under ~800 tokens

Auto-injection into new sessions

When a new session ID appears in the .jsonl files (i.e. you started a fresh Claude conversation), the extension automatically:

1. Reads MEMORY.md

2. Copies it to your clipboard

3. Shows a notification:

   "Claude memory loaded (~320 tokens) — paste at the start of your prompt with Ctrl+V to give Claude full project context"

This avoids any monkey-patching of VS Code internals. Just Ctrl+V at the start of your prompt.

Commands

Bug Fix Search

Run Claude Token Manager: Search Bug Fixes to open a searchable list of every bug fix ever recorded. Search by:

• Bug title / description
• Tags (e.g. typescript, auth, ui)
• File names

Selecting an entry opens a detail panel with the full bug report. Use this before asking Claude to solve a problem you may have already solved.

Privacy & Security

• The extension only reads .jsonl files written by Claude Code. It never writes to them.
• Your API key is stored in VS Code SecretStorage (OS-level keychain on Windows/macOS, libsecret on Linux). It is never written to disk in plain text.
• The API key is used only for the summarization call to claude-haiku-4-5-20251001. No other data is sent to Anthropic beyond what you explicitly summarize.
• claude-memory/ files are local to your workspace. Add claude-memory/ to .gitignore if you don't want them committed.

How Claude Code JSONL files are located

Claude Code stores session transcripts in:

~/.claude/projects/{encoded-workspace-path}/*.jsonl

The workspace path encoding replaces path separators with hyphens. For example:

The extension tries several encoding variants automatically to handle platform differences.

Configuration

One optional setting is available in VS Code settings (Ctrl+,):

Limitations

• The token panel opens with a snapshot at click time and does not auto-refresh. Close and reopen to see updated data, or use the status bar for live numbers.
• If Claude Code creates a .jsonl file and adds the session ID before the extension detects it as "new", the clipboard injection may be missed for that session. Subsequent sessions will always be caught.
• Summarization requires an internet connection and a valid Anthropic API key. If summarization fails the session data is not lost — you can retry with Summarize Session Now.
• The cost estimate uses Sonnet 4.5 pricing regardless of the actual model Claude Code uses. Check costUSD fields in the .jsonl files for exact costs.

Author

PowerisTsutusn