Token Guardian

Token Guardian is a CLI-first guardrail for checking prompt size, context pressure, and estimated cost before you call an LLM.

It helps developers answer three questions quickly:

• how many tokens this prompt will probably use
• how much this request may cost
• whether this prompt is risky for the selected context window

Preview

The current CLI interface is optimized for interactive terminal use and the guided experience is presented in pt-BR.

Why Use Token Guardian?

• catch oversized prompts before they hit the model
• estimate cost before expensive runs
• compare supported models using the same prompt
• clean duplicated or bloated prompt text
• keep simple local observability with SQLite metrics
• start from an interactive terminal menu instead of memorizing commands

Copy-Paste Install

Windows PowerShell

pip install token-guardian
token-guardian

macOS / Linux

pip install token-guardian
token-guardian

Running token-guardian without arguments opens an interactive menu when your terminal supports it.

What the CLI Can Do

• analyze one prompt for one provider/model pair
• compare one prompt across the default supported models
• optimize prompt text by removing duplicates and excess whitespace
• list supported models
• sync the local model catalog snapshot
• inspect local usage metrics
• guide the full flow through an interactive menu

Usage

Start here

token-guardian

Running without arguments shows the available flow and the most useful commands to start with.

In interactive terminals, Token Guardian opens a guided menu with:

• provider selection
• model selection
• prompt entry with Enter to send
• sync selection by provider
• quick access to models and metrics

Analyze a prompt

token-guardian analyze \
  --provider anthropic \
  --model claude-sonnet-4 \
  --prompt "Review this architecture proposal and identify risks."

Compare models

token-guardian compare \
  --prompt "Summarize this technical RFC and list migration risks."

Optimize a prompt

token-guardian optimize \
  --prompt "Goal: summarize
Goal: summarize


Return bullets only."

List supported models

token-guardian models

Sync model catalog

token-guardian sync-models
token-guardian sync-models --provider openai

View local metrics

token-guardian metrics

Example Output

Typical analyze output is rendered as a terminal report with token estimate, cost, context usage, risk, and prompt guidance.

Supported Providers

Current built-in catalog:

• OpenAI: gpt-4.1
• Anthropic: claude-sonnet-4, claude-opus-4
• Google: gemini-2.5-pro, gemini-2.5-flash
• OpenRouter: openai/gpt-4.1

Each model stores:

• context limit
• input price per 1K tokens
• output price per 1K tokens
• speed estimate
• source URL

The CLI also shows catalog metadata such as:

• Catalogo atualizado em 2026-06-13
• the current JSON snapshot path

Scoring

Risk level

Based on estimated context usage:

• low
• medium
• high
• critical

Context health score

Range: 0 to 100

Factors include:

• prompt size
• repeated lines
• repeated vocabulary
• redundant sections

Cost score

• $: very low
• $$: low
• $$$: medium
• $$$$: high

Complexity score

• Simple
• Medium
• Complex
• Very Complex

Metrics

Token Guardian stores local metrics in SQLite.

Database file:

• token_guardian.db

Tracked fields include:

• total requests
• total tokens
• estimated cumulative cost
• top models
• top providers

Project Structure

token-guardian/
|-- app/
|   |-- cli.py
|   |-- models/
|   |-- providers/
|   |-- services/
|   `-- utils/
|-- docs/
|-- tests/
|-- LICENSE
|-- pyproject.toml
`-- README.md

Development

Run tests:

pytest

Run quality checks:

ruff check .
black --check .
mypy app

Roadmap

• add richer interactive CLI flows
• expand supported model catalog
• improve prompt optimization heuristics
• add exportable reports
• add model catalog sync support

License

MIT