Live now on Cloudflare Pages and Workers

Know the cost before Claude Code runs.

Candura predicts prompt cost, tracks live spend across sessions, and gives you a practical way to keep AI-assisted coding honest. It is built to launch cheaply, explain itself clearly, and stay easy to move to a custom domain later.

View on GitHub npm package See research basis

Predict

Classifies the task, estimates how much context Claude Code will need, and surfaces a cost range before execution.

Track

Follows live spend across active sessions so long-running work does not drift past budget unnoticed.

Improve

Anonymous usage data helps refine heuristics from real prompts instead of guessed averages.

How it works

Candura is built around a simple loop: estimate first, observe while running, then calibrate from reality. That gives you useful defaults for solo work and a path to better accuracy over time.

Execution flow

Before each prompt, Candura classifies the task and predicts likely cost.
During the run, it monitors live spend so you can see escalation early.
After completion, it records the actual outcome and feeds that back into future estimates.
For larger work, it can suggest smaller or more direct prompts that reduce context churn.

What ships today

Root package candura

Workspace @candura/core, @candura/hooks, @candura/daemon

Site candura.pages.dev

Install and launch

The public launch path is intentionally lightweight: clone, bootstrap, and you are ready to use the package and local tooling.

Install

git clone https://github.com/Madhav2310/PredictorAgent.git
cd PredictorAgent && ./setup.sh

The daemon starts automatically on first prompt, so there is no extra terminal chore to manage.

Useful commands

candura predict "your prompt"
candura stats
candura history
npm run tui

The TUI is especially useful when you are juggling multiple Claude Code windows at once.

Telemetry and privacy

Telemetry exists to make the heuristics sharper, not to collect more than is needed. The defaults are transparent, and the opt-out path stays simple.

Configuration

{
  "thresholds": {
    "warn_above": 2.0,
    "block_above": 6.0,
    "daily_budget": 50.0
  },
  "telemetry": {
    "enabled": true,
    "endpoint": "https://candura.candura-telemetry.workers.dev/v1/session"
  }
}

What is sent

Prompt text, predicted cost, actual cost, task type, model, and run duration.
No file contents, code diffs, file paths, or machine identifiers are included.
Telemetry can be disabled in ~/.candura/config.json if needed.
The endpoint can also be overridden with CANDURA_TELEMETRY_ENDPOINT.

Research basis

The public site now reflects the fact that Candura is not just a wrapper around a heuristic. It is grounded in measurement, and the launch copy should say that plainly.

Evidence base

Measured against real Claude Code sessions and API response patterns.
Calibrated using cost signals that matter in practice, especially context growth and file reads.
Inspired by published work on agentic coding efficiency, token economics, and prompt optimization.
Designed to remain honest about uncertainty instead of hiding variance behind a single number.

Live endpoints

Site candura.pages.dev

Telemetry workers.dev

Package npmjs.com/package/candura

Why this version exists

It gives the project a real public face instead of a placeholder landing page.
It makes the package, telemetry, and research story visible in one place.
It keeps the launch cheap today and flexible for a future custom domain.
It tells users what is happening with enough specificity to build trust quickly.