# Tokalator

> Open-source context engineering for AI coding assistants (AGPL-3.0-or-later). Living code maps for agents (GLOSSARY.md and SITE-MAP.md), a VS Code extension that tracks context budgets and usage across Claude Code, GitHub Copilot and Codex, an MCP server and CLI, and browser-based token and pricing tools for 32 Claude, GPT, and Gemini models. Updated 2026-10-10.

Tokalator helps developers see what fills the context window, what a session costs, and which files matter. Everything runs locally or in the browser. Telemetry in the extension is opt-in and off by default.

## VS Code extension (4.2.0)

- [Extension overview, live dashboard demo, settings, and commands](https://tokalator.ai/extension): context budget meter, tab relevance scoring and cleanup, pinned files, Turn Analytics, input-only turn cost with long-context pricing.
- Agent usage, read locally: Claude Code (CLI and VS Code extension session logs), GitHub Copilot Chat (the session files VS Code stores for the workspace, with token counts where recorded and estimates otherwise), GitHub Copilot CLI, and OpenAI Codex CLI. Shows turns, requests, cache hit rate, context size per turn, and cost.
- 4.2.0 fixes: turn counts no longer stay at 0; the budget meter, Turn Analytics and Context Growth follow the agent you use; only sessions from the open workspace are counted; a model you pick stays for 30 minutes and does not change other VS Code windows.
- 4.2.0 additions:
  - All tab: Claude Code, Copilot Chat and Codex side by side (today's usage, turns, context used, model).
  - Agent memory and instructions: the CLAUDE.md files, Claude Code memory, skills, subagents, AGENTS.md and Copilot instructions each agent loads, with token counts; click to edit.
  - Pin into your agent: pinning keeps a file at full relevance and adds it to the next message of Claude Code (@-mention), Copilot Chat or Codex.
  - Session facts measured from the logs: context used, turns left, subagent share, output including thinking, cache share, tokens in the last 5 hours.
  - Workspace tools by click: Agent Instructions (AGENTS.md from project facts plus a CLAUDE.md that imports it), Checks Skill (a SKILL.md that runs type check, lint, tests and build, in .agents/skills and .claude/skills), Secure Workspace, Code Maps, Price Watch (rate changes priced on each agent's last 7 days).
  - Status bar shows the followed agent's real context; Get Started walkthrough; Open Dashboard in Editor.
- 4.2.0 beta: Dependency Check (known vulnerabilities from OSV.dev in every lockfile, outdated direct dependencies, runtime end-of-life from endoflife.date), Claude Code hooks (list, best-practice checks, open, remove, pause), and What Claude Code runs (tools, command-line program names, MCP servers, last 7 days).
- @tokalator chat commands: /count, /breakdown, /optimize, /pin, /unpin, /instructions, /model, /compaction, /preview, /secure, /skill-gen, /settings-gen, /codemap, /pricewatch, /reset, /exit.
- API-equivalent cost means what the tokens would cost at API list prices; Claude, Copilot and ChatGPT subscriptions are not billed per token.
- [Marketplace listing](https://marketplace.visualstudio.com/items?itemName=vfaraji89.tokalator)

## Tools (run in the browser)

- [Claude Code usage](https://tokalator.ai/tools/usage): drop ~/.claude/projects logs to see turns, requests, cache hits, and API-equivalent cost by model, project, client, and day. Files are not uploaded.
- [Cost calculator](https://tokalator.ai/tools/calculator): compare request cost across all catalog models, with cache and long-context rates.
- [Model comparison](https://tokalator.ai/tools/compare) and [pricing reference](https://tokalator.ai/tools/pricing): rates per million tokens, context windows, and provider sources.
- [Prompt caching](https://tokalator.ai/tools/caching), [conversation growth](https://tokalator.ai/tools/conversation), [context budget](https://tokalator.ai/tools/context), [economics](https://tokalator.ai/tools/economics).

## Learn

- [Learn](https://tokalator.ai/learn): lessons on context windows, server-side compaction, context editing, prompt caching, effort, and tool search, with sources.
- [Dictionary](https://tokalator.ai/dictionary): 99 context engineering and AI coding terms, each with a plain-language line, a technical definition, an example and related terms.
- [Benchmarks](https://tokalator.ai/benchmarks): what SWE-bench, Terminal-Bench, Humanity's Last Exam, GPQA, ARC-AGI and other benchmarks test, how they are scored, and the published scores of current models.
- [Wiki](https://tokalator.ai/wiki): arXiv papers on long context, context rot, prompt caching, and agent context files.

## MCP server and CLI

- [tokalator-mcp](https://tokalator.ai/extension#mcp): MCP tools count_tokens, estimate_budget, preview_turn, list_models, and claude_code_usage for Claude Code; the same as a terminal CLI (tokalator count, budget, preview, models, usage --watch).

## Project

- [Source on GitHub](https://github.com/vfaraji89/tokalator)
- [Full text for LLMs](https://tokalator.ai/llms-full.txt)
- [Privacy policy](https://tokalator.ai/privacy)
- [Paper: Tokalator, a context engineering toolkit for AI coding assistants (arXiv:2604.08290)](https://arxiv.org/abs/2604.08290)