Visual Vortex
Visual Vortex
Open-source tools for AI in production

AI in production,without the surprises.

Visual Vortex builds open-source developer tools for running AI in production. TokenSense shows what every LLM call costs and routes it to the right model; LoopBrake stops agent tasks that run far past their normal length.

See our products GitHub
TokenSenseAI Cost Intelligence
LoopBrakeRunaway Agent Control

Works with the tools you already use

  • OpenAI
  • Anthropic
  • Groq
  • Claude Code
  • Codex
  • Python
Our products

Two tools.
No surprises.

Open source, MIT licensed, and running inside your own stack.

Runaway Agent Control

LoopBrake

LoopBrake learns how long your agent's tasks normally take, stops the one that runs way past that, and tells you why. For Claude Code, Codex and your own Python agents.

pip install loopbrakeExplore LoopBrake

AI Cost Intelligence

TokenSense

TokenSense tracks the tokens, cost and latency of every LLM call with one wrapper around your client, and routes requests across models by your rules. It runs in your process, so prompts never leave it.

pip install tokensense-aiExplore TokenSense

TokenSense

Every token matters.

Wrap your existing OpenAI, Anthropic or Groq client once, and every LLM call is tracked: tokens, cost and latency. It runs inside your process and records after the response returns, so calls stay as fast as before.

  • Every call, priced. Tokens, cost and latency for each LLM call, tagged by user, session or your own labels.

  • Smart routing. Context-budget-aware model switching, with automatic escalation when a call fails.

  • Your data stays yours. Prompts never leave your process.

Explore TokenSense Read the docs

Every call, priced

Example from the README

claude-sonnet-4-6

$0.0003

Tokens in
12
Tokens out
24
Latency
847 ms

Send the numbers to

Your terminalSQLiteYour loggerAny HTTP endpoint

What a stop looks like

Real public run

SWE-bench · GPT-5-mini

Step 27 of 72

Stopped at27 steps
Recorded run72 steps

Reason

Repeating in 5 of last 5 steps; nothing new in 2 of last 5 steps.

LoopBrake

Brakes for stuck AI agents.

LoopBrake learns how long your agent's tasks normally take and stops the one that runs way past that, with the reason in plain words. Fewer than 1 in 20 good tasks get stopped, as long as your future tasks are like your past ones.

  • Learns your normal. Each agent's limit comes from its own past tasks. No step count to guess.

  • A promise it keeps. On 2,979 public agent runs, it stopped at most 4.8% of good runs on average.

  • Claude Code, Codex, your own agents. Plugins for both, and a Python package with a brake per agent.

Explore LoopBrake Read the docs

What we're exploring next

A daily call for parents who live far away.

An AI that calls elderly parents every day in their own language, reminds them about their medicines, and sends the family a short update each evening. We haven't built it yet. If your parents live far from you, tell us whether it would help.

Take the 2-minute survey

Every evening, on WhatsApp

Example · not a real family

Papa · today's updateWHATSAPP · 7:30 PM

  • Took his morning BP medicine
  • Mild knee pain, better than yesterday
  • Good mood · 9 pm reminder set

Start building with Visual Vortex.

Explore TokenSense →Explore LoopBrake →