ContextFlame

A context profiling tool for Claude Code sessions. Detects and visualizes context window bloat from tool calls — file reads, grep results, MCP responses — and shows you exactly where your tokens are going.

The Problem

Claude Code injects tool outputs into the conversation context. Over a session, these accumulate and cause context overflow, truncation, repeated file injection, and wasted tokens. There's no visibility into what's consuming the context window or when resets occur.

ContextFlame answers: where are my tokens going?

How It Works

ContextFlame wraps your claude command with a local proxy, like perf record or strace. It intercepts every API call, attributes token usage by source, and generates a flamegraph when you're done.

contextflame claude → proxy (auto port) → api.anthropic.com

No config. No background processes. When Claude exits, the proxy dies and you get your report.

Quick Start

pip install contextflame

Profile a session

contextflame claude

That's it. Use Claude Code as normal. When you exit, ContextFlame generates contextflame-report.html and prints a summary:

ContextFlame profiling → claude
Proxy on :52431 | Log: contextflame-1709571234.jsonl
────────────────────────────────────────────────────
  ... your Claude Code session ...
────────────────────────────────────────────────────
Report: /home/you/contextflame-report.html
  47 calls | 1,234,567 input tokens | 42% tool | 18% duplicate

Open contextflame-report.html in a browser.

Options

contextflame --port 8011 claude          # use a specific port
contextflame claude --model opus          # pass args through to claude
contextflame claude --resume              # works with any claude flags
contextflame --no-report claude           # skip report generation

Everything after contextflame [options] is passed directly to the wrapped command.

CLI Reference

contextflame <command> [args...]

Profile any command. Starts a proxy, runs the command with ANTHROPIC_BASE_URL pointed at it, generates a report when it exits.

Flag	Default	Description
--port	auto	Proxy port (0 = pick a free port)
--log	auto (timestamped)	Path to the JSONL log file
--output	contextflame-report.html	Report output path
--no-report	off	Skip report generation

contextflame report

Generate a report from an existing log file.

Flag	Default	Description
--log	(required)	Path to the JSONL log file
--output	contextflame-report.html	Output HTML file path

contextflame watch

Live-tail token usage in the terminal with sparkline bars.

Flag	Default	Description
--log	(required)	Path to the JSONL log file
--interval	1.0	Refresh interval in seconds

ContextFlame watch — press Ctrl+C to stop
────────────────────────────────────────────────────────────────────────
#a1b2c3d4 ▃▃▃▃▃▃░░░░░░░░░░░░░░ in: 32.1k out: 1.2k tool: 18.4k  Read×3 Grep×1
#e5f6a7b8 ▄▄▄▄▄▄▄▄░░░░░░░░░░░░ in: 45.8k out:  842  tool: 22.1k  Read×2 Bash×1
#c9d0e1f2 ▆▆▆▆▆▆▆▆▆▆▆░░░░░░░░░ in: 98.3k out: 2.1k tool: 51.2k  Read×5 Grep×2 Bash×1

contextflame start

Run the proxy server standalone (for advanced use or if you want separate terminals).

Flag	Default	Description
--port	8011	Port to run the proxy on
--host	0.0.0.0	Host to bind to
--log	contextflame.jsonl	Path to the JSONL log file

The Report

The HTML report is a self-contained file with:

• Metrics bar — total calls, input/output tokens, tool token ratio, duplicate ratio, peak utilization, context resets, wasted tokens
• Flamegraph — width proportional to tokens consumed, click to zoom in:
  • Session → API calls → categories (System/History/Tools/Output) → individual tool results
  • Breadcrumb navigation to zoom back out
  • Duplicate content shown with red stripe pattern
• Token timeline — stacked bar chart showing context growth over the session
• Top tables — tools and files ranked by cumulative token consumption

What It Tracks

Metric	What it tells you
Tool token ratio	What fraction of your context is tool outputs vs. actual conversation
Duplicate ratio	How often the same file/content gets re-injected across calls
Top tools	Which tools (Read, Grep, Bash, MCP) consume the most tokens
Top files	Which files get read most often and cost the most tokens
Context resets	When Claude Code's context window overflows and gets truncated
Peak utilization	How close you got to the 200k token limit
Wasted tokens	Tokens spent on duplicate content that was already in context

Architecture

src/contextflame/
├── cli.py           # Click CLI — wrap command, start, report, watch
├── proxy.py         # Starlette async proxy, handles streaming + non-streaming
├── attributor.py    # Parses request/response, classifies tokens by source
├── storage.py       # ContextSnapshot/ToolInjection dataclasses, JSONL I/O
├── metrics.py       # Computes session-level bloat metrics
├── flamegraph.py    # Renders the HTML report from snapshots
└── templates/
    └── flamegraph.html  # D3.js interactive flamegraph template

Development

uv run pytest tests/ -v