⚡️NEW: tokmon beam - a self-hosted web UI to monitor API usage data, running from localhost.

 ______   ______    __  __    __    __    ______    __   __    
/\__  _\ /\  __ \  /\ \/ /   /\ "-./  \  /\  __ \  /\ "-.\ \   
\/_/\ \/ \ \ \/\ \ \ \  _"-. \ \ \-./\ \ \ \ \/\ \ \ \ \-.  \  
   \ \_\  \ \_____\ \ \_\ \_\ \ \_\ \ \_\ \ \_____\ \ \_\\"\_\ 
    \/_/   \/_____/  \/_/\/_/  \/_/  \/_/  \/_____/  \/_/ \/_/ 
                                                                   

CLI utility to monitor your OpenAI API token usage

tokmon 🔤🧐 - CLI to monitor OpenAI API usage

tokmon (Token Monitor) lets you keep track of your program's OpenAI API token usage.

You can use tokmon just like you would use the time utility, but instead of execution time you get OpenAI usage and cost stats.

Installation

pip install tokmon

Usage

Note: tokmon works for gpt-* models (gpt-3.5-turbo, gpt-4, etc.). If you need support for other models (e.g. davinci) see tokmon#6.

$ tokmon /path/to/your/<your program> [arg1] [arg2] ...

tokmon cost report:
================================================================================
Monitored invocation: python3 ./tests/python_example.py --prompt say 'hello, tokmon!'
Models: ['gpt-3.5-turbo-0301']
Total Usage: {'total_prompt_tokens': 26, 'total_completion_tokens': 12, 'total_tokens': 38}
Pricing: {'gpt-3.5-turbo-0301': {'prompt_cost': 0.002, 'completion_cost': 0.002, 'per_tokens': 1000}}
Total Cost: $0.000076
================================================================================

Writing cost summary to JSON file: ./tokmon_usage_summary_1682039505.json

After your program finishes running (or you ctrl^C out it), tokmon will print a summary that looks like the above. tokmon also generates a detailed report and saves it as a JSON file.
You can use the --beam <url> flag to stream token usage data to a server. See tokmon --beam for more information.

Demo

Try it out

# Install tokmon
pip install tokmon

# Clone this repo and `cd` into it
git clone https://github.com/yagil/tokmon.git && cd tokmon/

# export your OpenAI API key. This will be used in the test program (source in ./tests/python_example.py)
export OPENAI_API_KEY="YOUR_OPENAI_API_KEY" 

# Run tokmon (save the JSON summary to the current folder)
# The test uses the openai python package. Install it if you don't have it: `pip install openai`
tokmon --json_out=. python3 ./tests/python_example.py --prompt "say 'hello, tokmon!'"

tokmon also supports this:

• If your program uses multiple OpenAI models in the same invocation, their respective usages will be reflected in the report.
• You can run multiple instances of tokmon simultaneously. Each invocation will generate a separate usage report.
• Pass a --json_out /your/path/report.json to get a detailed breakdown + conversation history in JSON format.

---

Use tokmon with your application or script

Warning This is a debugging tool. It is not intended to be used in any consequential setting. Use your best judgement!

Prepend tokmon to your normal program invocation like so:

$ tokmon /path/to/your/<your program> [arg1] [arg2] ...

Run and use your program just like you would normally (arguments and all). Interactive usage is supported as well.

Python

Simply add tokmon to the beginning of your regular invocation

$ tokmon python /path/to/your/script.py

This will work for scripts and long running programs like Django / Flask / FastAPI servers.

Node

For scripts:

$ tokmon node /path/to/your/script.js

npm run

Edit your package.json's "scripts" entry to include tokmon.

{
    "scripts": {
        "dev": "tokmon next dev",
        ...
    }
}

Full usage and cost summary (JSON)

{
    "total_cost": 0.0019199999999999998,
    "total_usage": {
        "total_prompt_tokens": 18,
        "total_completion_tokens": 23,
        "total_tokens": 41
    },
    "pricing_data": "{'gpt-4-0314': {'prompt_cost': 0.03, 'completion_cost': 0.06, 'per_tokens': 1000}}",
    "models": [
        "gpt-4-0314"
    ],
    "raw_data": [
        {
            "model": "gpt-4-0314",
            "usage": {
                "prompt_tokens": 18,
                "completion_tokens": 23,
                "total_tokens": 41
            },
            "cost": 0.0019199999999999998,
            "messages": [
                {
                    "role": "system",
                    "content": "You're a helpful assistant."
                },
                {
                    "role": "user",
                    "content": "hello"
                },
                {
                    "role": "assistant",
                    "content": "Hello! How can I help you today? If you have any questions or need assistance, feel free to ask."
                }
            ]
        }
    ]
}

How it works

tokmon uses the mitmproxy library to intercept HTTP requests and responses between your program and the OpenAI API. It then processes the request and response data to calculate token usage and cost based on tokmon/openai-pricing.json.

tokmon works for programs in python / node (using OpenAI's clients), or curl (run directly, and not i.e. in a bash script). See Golang instructions for instructions on how to use tokmon with Golang programs.

if you manually install mitmproxy's CA certificate, it should work for all executables (note: haven't tested this.)

In most cases, tokmon relies on the 'usage' field in OpenAI's API responses for token counts. For streaming requests, however, tokmon uses OpenAI's tiktoken library directly to count the tokens. As of writing OpenAI's API does not return usage data for streaming requests (reference.)

openai-pricing.json

The pricing data was extracted from OpenAI's website with the help of ChatGPT.

tokmon is using tokmon/openai-pricing.json from its package.

{   
    "last_updated": "2023-04-12",
    "data_sources": [
        "https://openai.com/pricing",
        "https://platform.openai.com/docs/models/model-endpoint-compatibility"
    ],
    "gpt-4": {"prompt_cost": 0.03, "completion_cost": 0.06, "per_tokens": 1000},
    "gpt-4-0314": {"prompt_cost": 0.03, "completion_cost": 0.06, "per_tokens": 1000},
    "gpt-4-32k": {"prompt_cost": 0.06, "completion_cost": 0.12, "per_tokens": 1000},
    "gpt-4-32k-0314": {"prompt_cost": 0.06, "completion_cost": 0.12, "per_tokens": 1000},
    "gpt-3.5-turbo": {"prompt_cost": 0.002, "completion_cost": 0.002, "per_tokens": 1000},
    "gpt-3.5-turbo-0301": {"prompt_cost": 0.002, "completion_cost": 0.002, "per_tokens": 1000},
    "text-davinci-003": {"cost": 0.02, "per_tokens": 1000},
    "text-curie-001": {"cost": 0.002, "per_tokens": 1000},
    "text-babbage-001": {"cost": 0.0005, "per_tokens": 1000},
    "text-ada-001": {"cost": 0.0004, "per_tokens": 1000},
    "text-embedding-ada-002": {"cost": 0.0004, "per_tokens": 1000}
}

You can override the default pricing with: tokmon --pricing /path/to/your/custom-openai-pricing.json ...

This pricing JSON is incomplete (missing DALL-E, etc.), it may be incorrect, and it may go out of date.

For best results, make sure to check that you have the latest pricing.

Misc

Golang

Hacky workaround for Golang programs. Add this to your program:

// Import these packages
import (
	"os"
	"crypto/tls"
	"crypto/x509"
	"io/ioutil"
	"net/http"
)

// Place this code somewhere in your go program *before* you make any calls to OpenAI's API.
certFile := os.Getenv("TOKMON_SSL_CERT_FILE") // This env variable will be set by tokmon
caCert, err := ioutil.ReadFile(certFile)
if err == nil { 
	caCertPool, _ := x509.SystemCertPool()
	caCertPool.AppendCertsFromPEM(caCert)
	http.DefaultTransport.(*http.Transport).TLSClientConfig = &tls.Config{
		RootCAs: caCertPool,
	}
}

Current Limitations

1. Event streaming: tokmon buffers Server-Sent Events (SSE) until the data: [DONE] chunk is received. If the monitored program leverages event streaming, its behavior will be modified.
   • Issue: tokmon#4
2. Only chat models are supported (gpt-3.5-turbo, gpt-4 and variants)
   • Issue: tokmon#6

Contributing

If you'd like to contribute to the project, please follow these steps:

1. Fork the repository.
2. Create a new branch for your changes.
3. Make your changes and test them.
4. Submit a pull request with a clear description of your changes and any relevant information.

Warning

1. tokmon comes without any warranty or guarantee whatsoever.
2. tokmon was tested on macOS only. It might not work on other platforms.
3. This tool may not work as intended, have unknown side effects, may output incorrect information, or not work at all.
4. The pricing data in openai-pricing.json may go out of date.