contextmeter 0.1.0

A live meter for your LLM's context window — see exactly what's filling it.

contextmeter

A live meter for your LLM's context window — see exactly what's filling it.

contextmeter is a tiny local proxy + terminal HUD. Point your LLM app at it and watch, in real time, how your context window fills up — broken down by system prompt, tool definitions, chat history, and tool outputs. It makes the invisible "tool tax" obvious: "34% of your window is tool schemas you never call."

It auto-detects the request format per call, so it works with any OpenAI-compatible API and Anthropic out of the box — no code changes, just a base-URL swap.

Works with

Anything that speaks the OpenAI chat format — OpenAI, OpenRouter, Groq, Together, Mistral, DeepSeek, Fireworks, and local servers like Ollama, vLLM, and LM Studio — plus Anthropic (Claude). contextmeter figures out which format each request uses on its own.

Install

uv pip install contextmeter

Use

Start the meter (provider is auto-detected; defaults to forwarding to OpenAI):

contextmeter

Then point your app's base URL at http://localhost:8765. To forward somewhere else (Anthropic, a local server, OpenRouter, …), just set --upstream:

contextmeter --upstream https://api.anthropic.com
contextmeter --upstream http://localhost:11434

--provider is optional — pass openai or anthropic only if you want to pin the format instead of letting it auto-detect.

Note: pass the host root as --upstream, not the /v1 endpoint. Your LLM client already appends the path (e.g. /v1/chat/completions), so use http://localhost:11434 — not http://localhost:11434/v1 (which would double the path to /v1/v1/...).

Why

Most developers have no idea what's eating their context window. Tool definitions alone can consume 40–60% before the first message. contextmeter shows you, live, in your terminal.

Status

Early. Supports OpenAI-compatible + Anthropic request analysis with automatic per-request provider detection. Token counts are approximate (tiktoken-based; Anthropic is approximated). Responses are currently buffered (no token-by-token streaming pass-through yet), and the breakdown counts message text but not yet the model's emitted tool_calls/tool_use argument payloads.

Roadmap: native Google Gemini schema; streaming pass-through; count tool-call argument payloads in History; Python wrapper entry point (no proxy); pricing/limits auto-refresh; richer Textual UI.