Overview / AI Spend Live

Stop guessing which agent session burned the budget.

AI Spend Live is a local dashboard for Claude Code and Codex usage logs. It shows the costly sessions, giant turns, cache behavior, and provider split without uploading your private prompts.

"You are not out of ideas. You are out of context budget."

Run the local dashboard View sanitized sample npm run dashboard

Sanitized dashboard from real local usage.

The numbers below preserve the shape of a real local 7-day run. Session names are generic and hashes, paths, prompt text, and repo-specific labels are removed.

Selected range: 7 days / API-equivalent estimate

Selected range 7,526.3M $12,816.46 estimate

24h burn 153.6M $236.07 estimate

Last hour 15.3M current pace

Cache reuse 81% 7,328.1M cached/read tokens

Per Session Burn generic names only

Session	Provider	Duration	Cache	Tokens	Estimate
Long Context Refactorlargest repeated-context thread	Claude	6.2d	99%	3,310.4M	$5,536.19
UI Polish Sessionlarge ongoing design pass	Claude	1.1d	99%	1,244.8M	$2,378.11
Codex Repository Sweeplong-running codebase work	Codex	3.0d	49%	928.0M	$707.37
Feature Handoff Sessionsingle thread carrying context	Claude	5.4d	96%	709.1M	$1,920.07

Why this matters what the dashboard surfaces first

One Claude thread carried 44% of the burn. That is the session to split, summarize, or retire first.

High cache reuse did not mean low pressure. Cached/read tokens still dominated, so repeated context was the main driver.

The largest single turn hit 1.8M tokens. That is the clue for spikes that feel random during the workday.

Codex was cheaper per token, but still visible. Long Codex runs stood out instead of hiding behind Claude totals.

What it catches.

This is not a generic spend chart. It is a workbench for figuring out why coding agents suddenly feel expensive.

Session identity

Readable names before hashes.

Claude custom titles and Codex thread names are shown when available. Hashes stay secondary.

Limit pressure

Find long sessions that keep paying for context.

Duration, cache reuse, and token totals make stale high-burn sessions obvious.

Spike diagnosis

Largest turns explain sudden jumps.

Sort by turn size to find the exact moment the burn rate changed.

Provider split

Claude and Codex in the same view.

Compare model, provider, daily totals, and cache behavior without merging logs by hand.

Quick start.

Run it where Claude Code and Codex already store local session logs.

Clone the repo

Create the public GitHub repo from the prepared folder and clone it locally.

Start the dashboard

Launch the local parser and browser UI.

Use the range filters

Switch between 1 day, 7 days, 30 days, or provider-specific views to isolate spikes.

PowerShelllocal

npm install
npm run dashboard

# open http://127.0.0.1:9020/

Log paths.

Defaults are designed for a normal Windows install, but every source path can be overridden.

Default sourceslocal files

%USERPROFILE%\.claude\projects
%USERPROFILE%\.codex\sessions
%USERPROFILE%\.codex\session_index.jsonl

Overridesoptional

$env:CLAUDE_PROJECTS_ROOT="C:\path\to\.claude\projects"
$env:CODEX_SESSIONS_ROOT="C:\path\to\.codex\sessions"
$env:CODEX_SESSION_INDEX="C:\path\to\.codex\session_index.jsonl"
$env:AI_SPEND_PORT="9020"
npm run dashboard

Privacy model.

The public website is static. The real dashboard is local because your usage logs are private operational data.

100% local parsing

0 prompt text exposed

2 providers in v0.1

JSONL source format

Session names come from metadata only: Claude custom titles or agent names, Codex thread names, then project/session fallbacks.

Cost estimates.

Costs are API-equivalent estimates from local token logs and model pricing tables. They are not provider invoices, subscription limit counters, or guaranteed billing totals.

Deploy the public site.

Upload the static docs page to the web root for aispend.live. Do not deploy the local log parser publicly.

FTP upload folderstatic site

C:\devland\aispend-live\docs\