Your chosen files stay on this device

Understand your AI usage, or prepare your Obsidian vault for private AI memory.

Choose a Claude, ChatGPT, Codex, Cursor or GitHub Copilot history file for a usage report, or choose an Obsidian vault folder for a local 7C's readiness check. The files stay in this browser and are not sent to TOP.

Too much technical jargon? Not sure how this helps you?

Copy the prompt below and paste it into whatever AI assistant you already use. It asks the assistant to find out which tools you use and what you are actually worried about, then explain TOP in those terms.

The prompt is public information about TOP and nothing else. It carries no data from this page, it does not tell your assistant to look anything up, and it tells it not to ask you for your usage files.

Copies only the text shown below. Nothing is sent anywhere.

The exact prompt that gets copied

I am looking at a tool called TOP, the Token Optimisation Protocol, and I want to know in plain language whether it is any use to me.

Everything you need in order to answer is written below. Please do not open a web page, run a command, read a file on my computer, or search any connected app or service. This is a translation task, not a research task.

What TOP says it does today:
- It is a web page that runs entirely inside my own browser. I point it at the history files that my AI tools already write on my computer. The page reads those files locally and reports what that past usage cost, or what it would have cost at published pay-as-you-go API rates.
- It reads history from seven tools: Claude Code, Claude Chat, ChatGPT, Codex, Cursor, Cursor Composer and GitHub Copilot. It reports on one of them at a time, so if I use several I would run it once per tool and read the reports side by side.
- It states that my history files are never uploaded, and that the page makes no network request while it builds the report.
- It states that when it cannot match a model to a published rate, it shows the token count and marks the cost as unpriced, rather than showing it as zero.

What TOP does not do today:
- It describes past usage only. It does not state that any bill would have been different.
- It does not pick models for me and it does not personalise anything. Those are described as research directions, not as working features.
- It cannot see a provider's actual invoice. Some tools record a real charge, some record only token counts, and some usage is covered by a subscription with no per-use charge. A figure worked out from published rates is a comparison, not a bill, and the two should not be added together.

Please treat all of the above as TOP's own claims about itself rather than as independent proof. Where a claim would need checking before I relied on it, say so.

Rules for this conversation:
1. Please do not ask me to upload, paste, attach or show you any usage export, log, prompt, source code, report, invoice, account identifier, credential, API key or confidential material. You do not need any of it to answer this.
2. Please do not read my files or my other conversations in order to personalise your answer. Use only what I tell you here.
3. If you conclude that TOP is not useful to me, please say so plainly. I would rather hear that than a sales pitch.

Please begin by asking me these two short questions:
1. Which AI tools do I actually use, and roughly how heavily?
2. What is my real worry: that I am spending more than I think, that I cannot tell which tool is costing what, that I need a defensible number to show someone else, or something else?

I am free to answer briefly or to skip either question.

Then, using whatever I tell you, explain:
1. What TOP would and would not be able to show me, given the specific tools I named.
2. Whether my particular worry is one that TOP can address, one it can only partly address, or one it cannot address at all.
3. What its report would still be missing for me, including anything it would not be able to price.
4. One honest sentence on whether it is worth ten minutes of my time.

Keep it short and concrete, in ordinary language. If you have to use a technical term, define it in the same sentence.

Sunrise: tell us what you want to improve8%

1Which Of These Applies To You?

Choose every statement that sounds like you. This helps TOP show the most useful parts of your report first.

You can make your report now. TOP is not accepting files yet.Your original file stays on this device. Please do not send files to TOP while we finish clear rules for permission, storage and deletion.

4Choose Where Your History Lives

This page opens the file on your device. It does not send the file anywhere.

Choose The Whole Claude Code Folder TOP opens only the supported .jsonl history files inside it. Everything stays on this device.

The usual Claude Code folder address

%USERPROFILE%\.claude\projects

TOP will show the folder-picker instructions for this computer here. A website cannot type a private path into the picker for you.

Use One Or More Individual Files Instead

A giant combined raw-history file still contains your prompts and can overwhelm a browser tab. The folder button above is the safer default. For a reusable single file, use the content-free privacy option after this report.

Choose Your File Or Files

Claude Code history files (.jsonl)

Tokens are small pieces of text counted by AI services. They are a measure of usage, not a separate charge on a subscription plan.

Usage By AI Version

Text sent to the AI: What you and your editor gave it to read, counted where it had to read that text fresh rather than from a saved copy.
Text returned by the AI: What it wrote back to you, including any code it produced.
Previously saved text: added: Text stored in a saved copy so the same words do not have to be read from scratch next time.
Previously saved text: reused: Text read again from that saved copy instead of fresh. It is billed far more cheaply than fresh text.

A Simple Example Of Price And Quality · cost information, plus invented example values

The green lines are invented only to explain the idea. TOP does not calculate what your work is worth.

cost recorded in your file possible spend after changing AI versions (estimate) invented example of useful work (not calculated) invented example after paying the AI cost (not calculated)

How much simple work might you try with a cheaper AI version? middle ground

Cost recorded in your file

Possible cost after changes estimate

Illustrative cost difference estimate

Pay-as-you-go means each amount of AI use is billed separately. Estimates use provider prices checked on 15 Jul 2026 and may differ from your bill. Claude Max, Claude Pro and ChatGPT subscriptions are not charged using the dollar estimate shown here.

Advanced Pricing Details

Prices vary by exact AI version, provider, reused input, service tier and date. AI versions we cannot match remain unpriced.

The reused-input estimate assumes Anthropic's five-minute rate at 1.25 times normal input. A one-hour reuse write is 2 times normal input. Claude Sonnet 5 uses its introductory $2 input and $10 output rate through 31 Aug 2026.

4Remove The Conversation Before Making A Copy

This advanced Claude Code option creates a new file on your computer. It keeps AI version names, usage numbers, dates and times, new numbered sessions made on your computer, and Claude Code access settings. It leaves out prompts, replies, code, file contents and the internal ID numbers from your original file. Read the tool, run it, and inspect the new file before choosing it here.

Show The Advanced Claude Code Instructions

Advanced option for Claude Code

Choose Download Privacy Cleaner below. It saves strip_claude_logs.py.
On Windows, open PowerShell in that download folder and run py .\strip_claude_logs.py. On macOS or Linux, run python3 strip_claude_logs.py. It writes claude-usage-only.jsonl next to it.
Open that file and check it yourself. It should contain no prompts, replies, code, file contents or original internal ID numbers. It can include dates, AI version names, Claude Code access settings and new numbered sessions.
Drop that file into the box below to see it parsed locally on your device.
Download a CSV or JSON copy for yourself. TOP is not accepting external files yet.

Nothing here uploads or sends itself. Keep the downloaded copy on your device for now.

strip_claude_logs.py

#!/usr/bin/env python3
"""
strip_claude_logs.py

Reads every Claude Code session log under ~/.claude/projects and writes ONE
new file, claude-usage-only.jsonl, in the folder you run this from.

What goes IN the output file: timestamps, model IDs, token counts (input,
output, cache read, cache write), permission mode, and locally generated
session numbers.

What never gets read into the output file: prompt text, response text,
tool inputs and outputs, file contents, code, and raw session, message,
request, or event IDs. Raw IDs are used only in memory to de-duplicate
multi-part turns and map the same work session to a new sequential number.
The raw IDs are then discarded before the new file is written.

The script is deliberately readable. Read it before you run it.

Usage:
    py strip_claude_logs.py
        Content-free usage fields only. No project names.

    py strip_claude_logs.py --with-projects
        Same as above, plus a "project" field per line: the folder name
        Claude Code created for that project. This can reveal your local
        folder structure and any internal codenames, so it is off by
        default. Only use it if you're fine sharing that.
"""
import json
import os
import sys
import glob
import math
import re

INCLUDE_PROJECT = "--with-projects" in sys.argv
SRC = os.path.expanduser("~/.claude/projects")
OUT = "claude-usage-only.jsonl"

KEEP_TOP = {
    "timestamp", "permissionMode",
}
KEEP_USAGE = {
    "input_tokens", "output_tokens",
    "cache_read_input_tokens", "cache_creation_input_tokens",
}
SAFE_LABEL = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,159}$")
SAFE_TIMESTAMP = re.compile(
    r"^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d+)?(?:Z|[+-]\d{2}:\d{2})$"
)
MAX_SAFE_INTEGER = 9_007_199_254_740_991


def safe_label(value):
    if isinstance(value, str) and SAFE_LABEL.fullmatch(value):
        return value
    return None


def safe_timestamp(value):
    if isinstance(value, str) and SAFE_TIMESTAMP.fullmatch(value):
        return value
    return None


def safe_counter(value):
    if isinstance(value, bool) or not isinstance(value, (int, float)):
        return None
    if not math.isfinite(value) or value < 0 or value > MAX_SAFE_INTEGER:
        return None
    if not float(value).is_integer():
        return None
    return int(value)


def strip_line(raw, project_name, session_number, line_number):
    try:
        rec = json.loads(raw)
    except json.JSONDecodeError:
        return None
    if not isinstance(rec, dict):
        return None

    msg = rec.get("message")
    if not isinstance(msg, dict):
        return None
    usage = msg.get("usage")
    if not isinstance(usage, dict):
        return None

    safe_usage = {}
    for key in KEEP_USAGE:
        value = safe_counter(usage.get(key))
        if value is not None:
            safe_usage[key] = value
    if not any(value > 0 for value in safe_usage.values()):
        return None

    out = {}
    timestamp = safe_timestamp(rec.get("timestamp"))
    if timestamp:
        out["timestamp"] = timestamp
    permission_mode = safe_label(rec.get("permissionMode"))
    if permission_mode:
        out["permissionMode"] = permission_mode
    out["model"] = safe_label(msg.get("model")) or "unknown"
    out["usage"] = safe_usage

    if INCLUDE_PROJECT and project_name:
        out["project"] = safe_label(project_name) or "unnamed-project"

    raw_session_id = rec.get("sessionId")
    private_session_key = (
        ("session", str(raw_session_id))
        if raw_session_id
        else ("file", session_number)
    )

    message_id = msg.get("id")
    request_id = rec.get("requestId")
    event_id = rec.get("uuid")
    if message_id or request_id:
        private_key = private_session_key + (
            "call", str(message_id or ""), str(request_id or "")
        )
    elif event_id:
        private_key = private_session_key + ("event", str(event_id))
    else:
        private_key = private_session_key + (
            "line", session_number, line_number
        )
    return private_key, private_session_key, out


def merge_duplicate(existing, incoming):
    for key, value in incoming.get("usage", {}).items():
        old = existing["usage"].get(key)
        if isinstance(value, (int, float)) and isinstance(old, (int, float)):
            existing["usage"][key] = max(old, value)
        elif key not in existing["usage"]:
            existing["usage"][key] = value
    if incoming.get("model") and not existing.get("model"):
        existing["model"] = incoming["model"]


def main():
    if not os.path.isdir(SRC):
        print(f"No Claude Code logs found at {SRC}")
        return

    n_in = 0
    safe_calls = {}
    call_order = []
    session_aliases = {}
    pattern = os.path.join(SRC, "**", "*.jsonl")
    paths = sorted(glob.glob(pattern, recursive=True))
    for session_number, path in enumerate(paths, start=1):
        project_name = os.path.basename(os.path.dirname(path))
        with open(path, encoding="utf-8", errors="ignore") as f_in:
            for line_number, line in enumerate(f_in, start=1):
                line = line.strip()
                if not line:
                    continue
                n_in += 1
                stripped = strip_line(
                    line, project_name, session_number, line_number
                )
                if not stripped:
                    continue
                private_key, private_session_key, safe_row = stripped
                if private_session_key not in session_aliases:
                    session_aliases[private_session_key] = len(session_aliases) + 1
                safe_row["session_number"] = session_aliases[private_session_key]
                if private_key in safe_calls:
                    merge_duplicate(safe_calls[private_key], safe_row)
                else:
                    safe_calls[private_key] = safe_row
                    call_order.append(private_key)

    with open(OUT, "w", encoding="utf-8") as f_out:
        for private_key in call_order:
            f_out.write(json.dumps(safe_calls[private_key]) + "\n")

    print(f"Read {n_in} log lines, wrote {len(call_order)} safe calls to {OUT}")
    print("Open that file and check it before you use it. It should")
    print("contain no prompt text, no code, and no file contents,")
    print("and no raw session, message, request, or event IDs.")


if __name__ == "__main__":
    main()

What the new file keeps: AI version names, usage numbers, dates, new numbered sessions, Claude Code access settings and the number of AI replies. What it leaves out: your prompts, the assistant's replies, code, file contents and the original internal ID numbers. You can read both the tool and its output before choosing the new file here.

Drop your claude-usage-only.jsonl here, or click to choose

the cleaned file the tool above produced, opened on this device and never uploaded

3Choose Your Obsidian Vault Folder

Your vault is memory, not a bill.

Obsidian notes usually do not contain reliable token or cost numbers, so TOP will not invent them. This path checks the shape of your vault and looks for separate Claude or ChatGPT history files you deliberately stored inside it.

Choose My Vault Folder

TOP checks file names and folder structure first. It does not read your Markdown notes, settings or attachments.

Choose Individual History Files Use this if you already have conversations.json, one or more Claude Code JSONL files, Codex usage JSONL, or an Anthropic Console CSV.

Desktop first: folder selection works best in a current desktop browser. A manual file option remains available.

Looking for your AI usage numbers?

Opening Claude Code inside an Obsidian vault does not put its usage logs inside that vault. Choose the separate Claude Code folder below. TOP reads the JSONL session files locally and keeps only their usage fields for this report. Your prompts, code and notes stay on this device.

Copy the Claude Code history path below. TOP will show the folder-picker instructions for this computer here. Do not copy raw history into a synced Obsidian vault.

%USERPROFILE%\.claude\projects

5Add Context To Your Report, Optional

We can explain the numbers better if you tell us what kind of work you do. Your answers stay in this tab. Skip any question, or skip the whole step.

How often do you use your AI assistant?

What do you primarily use your AI assistant for? choose any

What Kinds Of AI Work Do You Do? choose any

Choose every example that sounds like work you do.

How Much Effort Do You Usually Ask The AI To Use?

Choose the label you normally select. Higher levels may take longer and use more of your allowance.

What Would Make TOP Most Useful? choose any

Which Kind of Account Do You Mainly Use?

Your selections stay in this tab and can be added to the copy you deliberately download. They are not sent automatically.

Final Step

Download Or Copy My Safe Report

Exact File Prepared For Download

Read this before downloading. The downloaded text file will match this preview.

Research-Safe JSON

This separate file keeps every privacy-safe usage aggregate TOP calculated and the exact pricing rates applied. It includes the current value-model scenario only when the report qualifies for that illustration. Any included scenario is clearly marked as unvalidated, not a forecast result.

TOP-1 boundary: this aggregate can support analyzer validation and broad calibration research. It cannot train a pooled k-nearest-neighbour forecast because it contains no task chronology, prompt fingerprints or per-task realised cost. That needs a separate privacy-reviewed task-level format and separate consent. More pooled data may improve evaluation, but TOP does not promise that it will improve accuracy.

Before sharing it: inspect the exact JSON below. It contains aggregate usage totals, model labels, source and collector metadata, pricing references, optional answer categories, and an illustrative value scenario only when eligible. It contains no prompts, replies, code, tool output, paths, filenames, project or account identifiers, email addresses, exact times or original IDs. Nothing is sent until you deliberately use a sharing action.

privacy.network_delivery: "none" records that this analyzer did not transmit the report. If you later share the downloaded file, review the destination in your device first.

Understand your AI usage, or prepare your Obsidian vault for private AI memory.

1Which AI do you use?

2Give TOP your usage data

Ask your coding agent to prepare the safe file

Choose the usage folder

Download your Cursor usage CSV from cursor.com

Download your Copilot usage report

1Which Of These Applies To You?

2Choose Where Your Work Lives

Which Claude history do you have?

3What Do You Want To Do With Your Downloaded History?

4Choose Where Your History Lives

Your report is ready

Which models drove your usage?

Where your tokens actually went

What TOP Can Measure Today

Could TOP-1 learn from your history?

Monthly usage

Session shapes

Save your safe report

Usage By AI Version

A Simple Example Of Price And Quality · cost information, plus invented example values

Shareable Summary usage totals and categories, without your conversation text

Keep A Content-Free Codex Log In Obsidian

4Remove The Conversation Before Making A Copy

Advanced option for Claude Code

Usage By AI Version

Download Your Cleaned Copy

3Choose Your Obsidian Vault Folder

Looking for your AI usage numbers?

That is not your Obsidian vault

No supported AI history files found

7C's memory building blocks

Turn your vault into private AI memory with 7C's

5Add Context To Your Report, Optional

1Which AI do you use?

2Give TOP your usage data

Ask your coding agent to prepare the safe file

Choose the usage folder

Download your Cursor usage CSV from cursor.com

Download your Copilot usage report

1Which Of These Applies To You?

2Choose Where Your Work Lives

Which Claude history do you have?

3What Do You Want To Do With Your Downloaded History?

4Choose Where Your History Lives

Save your safe report locally

Your report is ready

Which models drove your usage?

Where your tokens actually went

What TOP Can Measure Today

Could TOP-1 learn from your history?

Monthly usage

Session shapes

Save your safe report

Usage By AI Version

A Simple Example Of Price And Quality · cost information, plus invented example values

Shareable Summary usage totals and categories, without your conversation text

Keep A Content-Free Codex Log In Obsidian

4Remove The Conversation Before Making A Copy

Advanced option for Claude Code

Usage By AI Version

Download Your Cleaned Copy

3Choose Your Obsidian Vault Folder

Looking for your AI usage numbers?

That is not your Obsidian vault

No supported AI history files found

7C's memory building blocks

Turn your vault into private AI memory with 7C's

5Add Context To Your Report, Optional

Download Or Copy My Safe Report

Exact File Prepared For Download

Research-Safe JSON

In plain English, here is everything in that file