elitesystem.ai tool

Context Budget Planner

Plan an agent's context window like a budget. Allocate the window across every consumer that competes for it, watch the stacked bar, and find the overflow here instead of in production.

Window

Pick a model or type your own number. The number in the box always wins.

Convenience list, checked 2026-07-31. Not authoritative: context limits change and vary by deployment. Verify against your provider's documentation.

Only line items sized by pasted text use this. It is a transparent character-based heuristic, not a tokenizer. The true count varies by tokenizer and by language, and code, JSON, and non-Latin scripts tokenize very differently from English prose. To see real token boundaries use tokenizer-visualizer; for cost use prompt-cost-calculator.

Assumptions

These drive the checks below. Defaults are starting points, not standards. Change them to match your workload.

Checked against your reserved output block.

Checked against leftover headroom.

User message plus assistant reply plus tool traffic, averaged.

Rule of thumb, editable. No standard exists.

Line items

Each item can be sized three ways: paste the actual text and have it estimated, enter a token count directly, or enter a unit count times a per-unit size. The arithmetic for every item is shown under it.

    The budget

    Every consumer, the reserved output block, and whatever headroom is left. Anything past the limit marker is overflow.

    Nothing planned yet.

    0 limit

      Checks

      The failures that actually bite a practitioner, evaluated against your numbers.

        Turns until overflow

        Given the current fixed cost and your average tokens per turn, how many more exchanges fit before the window is exhausted.

        -

        Export

        Drop the plan into a design doc or check it into a repo next to the prompt it describes.

        Markdown

        JSON

        What this is, and what it is not

        Honest boundaries.

        • Token counts from pasted text are estimates. A character-per-token divisor is a heuristic, not a tokenizer. Treat estimated items as approximate and enter measured counts when you have them.
        • Context limits are editable for a reason. The preset list is a convenience checked on 2026-07-31, not an authority. Providers change limits and deployments differ.
        • Thresholds are yours. "More than half the window on retrieval" and the tool-definition percentage are rules of thumb this tool exposes as inputs, not published standards.
        • Nothing is sent anywhere. No network requests, no external dependencies, no analytics. Your prompts stay in the tab.