Local-first  ·  No telemetry  ·  No external API calls

Find recurring AI agent jobs that quietly waste tokens before the waste compounds.

Audit scheduled and repeated agent runs locally, with evidence before manual fixes.
TokenSave helps you find avoidable recurring agent waste in your own imported logs. It does not prove vendor overbilling.

View GitHub → Try sample audit →
Runs locally
No telemetry
No external API calls
No auto-remediation
Evidence-backed manual fixes
No upload. No account.

How it works

Step 1
Import
Drop your job export file. All processing stays on your device.
Step 2
Diagnose
TokenSave scans for waste patterns and failure signals.
Step 3
Evidence
Each finding includes exact job IDs, token counts, and context.
Step 4
Manual Fix
Review the evidence and apply fixes yourself.

Why AI agent costs become hard to see

Agent jobs can keep consuming tokens in the background before you notice. A scheduled cron, a recurring subagent, or a long-running automation may keep firing while you are away from the keyboard. Token counts shown by the provider are local estimates — the per-session dollar figure is computed from token usage and may differ from your actual bill. Multi-instance usage, MCP tool calls, plugin activity, and background skills all add to the total. The hard part is rarely the total spend; the hard part is finding the avoidable recurring burn inside it.

What TokenSave can actually audit

TokenSave reads a job export and its run logs locally, then surfaces evidence-backed findings. Cost is shown as approximate exposure, not a guaranteed saving. Findings are surfaced only when your import contains the underlying signal. Categories we can detect when evidence is present:

Duplicate recurring jobs
Failure and retry loops
Context growth across runs
Stale active jobs
Risky silent runs (low-output, high-token jobs)
Model and routing risk
Tool and MCP overhead when logs include those signals

From confusing usage to local evidence

Run a local audit, see evidence, then manually decide what to fix. TokenSave never touches your provider account, never calls an external LLM, and never writes back to your scheduler. Fix guidance is text only; you copy the command, you run it. No upload, no account, no auto-fix.

What this is

TokenSave is:

What this is not

TokenSave is not:
Cost figures shown are approximate exposure based on local token counts and the pricing table you supply. They are not a guaranteed saving and they are not proof that any provider overcharged you.