tolkin-cli

Privacy-first AI token analyzer CLI. Count, audit, and optimize prompt and MCP token costs across OpenAI, Anthropic, and Gemini. Everything runs fully local: no telemetry, no uploads, nothing leaves your machine.

Install

npx tolkin-cli --help
# or
bunx tolkin-cli --help
# or install globally (installs the `tolkin` command)
npm i -g tolkin-cli

The tolkin-cli package is a thin launcher exposing the tolkin command. The actual binary ships in a per-platform package (tolkin-darwin-arm64, tolkin-darwin-x64) pulled in automatically as an optional dependency.

Commands

| Command | What it does | | --- | --- | | tolkin count <FILE> | Count tokens in a file or stdin. --all compares providers. | | tolkin compare <FILE> | Side-by-side token counts across providers. | | tolkin viz <FILE> | Visualize token boundaries (count plus estimate band for Claude). | | tolkin audit <FILE> | Run the rules engine: ranked findings with savings estimates. | | tolkin redact <FILE> | Strip secrets from input. Runs before anything else. | | tolkin cost <FILE> | Estimate provider cost for an input. | | tolkin mcp <CONFIG> | Analyze an MCP config: tool-definition token cost and CLI-swap savings. | | tolkin drift <FILE> | Compare the same input across tokenizer versions (encoding drift). | | tolkin scan | Discover your local agent configs (Claude, Cursor, Codex, VS Code, Zed, and more) and report MCP token costs, CLI-swap savings, instruction-file weight, and shell secret hygiene. Read-only. | | tolkin project [DIR] | Repo-wide audit: walks a repository (gitignore-aware), splits agent-context weight by load profile (always vs on-invocation vs on-demand), ranks the heaviest skill/command/instruction files, flags secrets, and totals reclaimable tokens. --fail-on high for CI. |

Use - as the file argument to read stdin, and --json for machine-readable output where supported. Run tolkin <command> --help for full flags.

echo "hello world" | tolkin count -
tolkin audit prompt.md
tolkin mcp ~/.config/claude/claude_desktop_config.json
tolkin scan

Supported platforms

macOS arm64 (Apple Silicon) and macOS x64 (Intel). Linux and Windows builds are coming.

License

MIT