Tokalator
See every token your AI coding spends. Context budget plus Claude Code, GitHub Copilot and Codex usage, tab relevance and secret-file protection, in your VS Code sidebar.
Install · Try it in your browser · Changelog · Free and open source (AGPL-3.0)


Features
- Context budget. Tokens used against your model's window, split into files, instructions, conversation and output reserve. Shows turns left before the window fills.
- Agent usage. Claude Code, GitHub Copilot (Chat and CLI) and Codex, each in its own tab: turns, tokens, cache hits and cost, read from their local logs. Updates every 15 seconds. Nothing to set up.
- All agents side by side. The All tab compares every active agent: today's usage, turns, context used and model.
Tokalator: Compare Agents opens it.
- Agent memory. Every file your agents load for this workspace (CLAUDE.md, Claude Code memory, skills, subagents, AGENTS.md, Copilot instructions) with its token count. Click to open and edit.
- Session facts. Context used, turns left, subagent share, output including thinking, cache share and the last 5 hours of usage, measured from the logs.
- Tokens or dollars. On a subscription? Switch to tokens and every dollar amount disappears.
- Tab relevance and pins. Every open tab gets a score from 0 to 1. Close the noise in one click. Pin a file to keep it at full relevance and add it to the agent you follow (right-click a tab or file, or use the pin in the sidebar).
- 32 current models. Claude, GPT and Gemini with verified prices and long-context rates. Detects your Copilot chat model.
- Secure Workspace. Checks that every secret file is hidden from agents, fixes the gaps, then verifies.
- Agent instructions and checks skill. One click writes
AGENTS.md from your real commands and layout (Codex and Copilot read it), a CLAUDE.md that imports it for Claude Code, and a SKILL.md that runs your checks before an agent finishes. Existing files are never overwritten.
- Dependency Check (beta). Every lockfile in the workspace, any ecosystem: known vulnerabilities from OSV.dev, outdated direct dependencies, and runtime end-of-life dates. Safe updates run only when you pick one.
- Claude Code hooks and tool use (beta). See, check, pause and remove hooks across all settings scopes, and see which tools and command-line programs Claude Code ran.
- Code maps. Writes
GLOSSARY.md and SITE-MAP.md so agents find the right files first. Built locally, no model calls.
What it tracks
| Assistant |
What it sees |
Cost |
| Claude Code |
Turns, tokens, cache hits, context size per turn |
Exact tokens |
| GitHub Copilot Chat |
Turns, model, prompt and output tokens from the sessions VS Code keeps |
Exact where recorded, estimated otherwise |
| Copilot CLI and Codex CLI |
Turns, tokens and model from their session logs |
Exact tokens |
| Any assistant |
Open tabs, pinned files, instruction and memory files |
Estimate |
Costs are API-equivalent: what the tokens would cost at API list prices. Subscriptions (Claude, Copilot, ChatGPT) are not billed per token, so you can show tokens instead with tokalator.costDisplay. VS Code records only the final prompt size of multi-step Copilot agent requests, so Copilot input can be undercounted.
Chat commands
Type @tokalator in chat:
| Command |
What it does |
/count, /breakdown |
Token count and where it goes |
/optimize |
Close low-relevance tabs |
/pin, /unpin |
Keep a file at full relevance |
/model |
Show or switch the model |
/compaction, /preview |
Growth per turn and next-turn preview |
/instructions |
Instruction files and their token cost |
/secure |
Workspace security audit |
/codemap |
Generate code maps |
/settings-gen, /skill-gen |
AGENTS.md with a CLAUDE.md that imports it, and a checks skill for Codex, VS Code and Claude Code |
/pricewatch |
Price changes priced on your agents' last 7 days |
Privacy
- Zero data retention. Your code, prompts, file names and paths never leave your machine, so nothing about your work is stored anywhere.
- Runs locally. The only default network request downloads the public pricing catalog. Agent logs, memory files and hook settings are read on your machine and never uploaded.
- Dependency Check, when you run it, sends package names and versions to OSV.dev, the package registries and endoflife.date. Nothing goes to Tokalator.
- Telemetry is off by default. When enabled it sends anonymous daily counts only, never code, file names, paths or prompts. Show, export or delete it any time from the Command Palette.
- Details: tokalator.ai/privacy
Settings
| Setting |
Default |
|
tokalator.model |
claude-opus-5-5 |
Fallback model when none is detected |
tokalator.relevanceThreshold |
0.3 |
Score below which tabs are flagged |
tokalator.claudeCode.enabled |
true |
Show Claude Code usage |
tokalator.agents.copilot.enabled |
true |
Show Copilot Chat and Copilot CLI usage |
tokalator.agents.codex.enabled |
true |
Show Codex CLI usage |
tokalator.costDisplay |
dollars |
Show usage as dollars, tokens or both |
tokalator.telemetry.enabled |
false |
Opt-in anonymous usage counts |
All settings: tokalator.ai/extension.
Feedback
Run Tokalator: Send Feedback or open an issue.
License
AGPL-3.0-or-later. Commercial licences are available from the copyright holder. "Tokalator" and its logo are trademarks of the copyright holder.
| |