Know what every prompt costs.
Before the invoice does.
PromptLedger is a 5-minute tracking service that shows what each feature and prompt costs per day — with Slack budget alerts and optional hard caps. From $29/mo flat. No per-seat pricing.
LLM bills triple silently
You discover it in the invoice, not in a dashboard. Datadog tells you the HTTP status was 200 — it has no idea which feature burned $3,000. Diagnosis takes days because nothing answers 'which prompts cost what.'
Unmonitored agent loops can burn thousands per incident; at scale it's a margin killer that shows up in the P&L unannounced.
LangSmith: $39/seat/mo.
PromptLedger Starter: $29/mo flat — cost-first, with budget enforcement.
How it works
- Add one logging call
After each provider call, POST the model + token counts to /v1/log — a few lines in any language, no SDK required. (Pass-through proxy mode — point your client at us instead — is on the roadmap.) - Tag by feature
One header (X-Feature: support-bot) attributes every call. - See the money
Per-feature, per-prompt cost today, this week, this month — dollars first, not spans. - Get warned — or stopped
Slack alerts on spikes; hard caps (Scale+) refuse over-budget log records with a 429 your app treats as a stop signal.
Why PromptLedger
Cost-first, not trace-first
The dashboard leads with dollars, not spans. One question answered perfectly: what did each feature cost today, and what's anomalous.
Budget enforcement
Per-feature budgets with Slack alerts — and optional hard caps that refuse over-budget log records with a 429 instead of recording them. That's the piece the $79/mo tools lack.
Flat pricing by tracked spend
No per-seat meter. $29 tracks $25k/mo of LLM spend.
5-minute setup
One POST per provider call from your own code — a snippet in any language, no SDK. Pass-through proxy mode is on the roadmap.
Questions, answered honestly
- How accurate are the costs?
- They're estimates from public per-token pricing — provider invoices are the source of truth. We say so in every response.
- What happens when a hard cap is hit?
- PromptLedger rejects the log record with a 429 and doesn't record it. The 429 answers your logging call — after that provider call has completed — so have your app stop calling the provider when it sees a 429 and the loop halts before its next call. Blocking inside the call path itself arrives with the pass-through proxy (roadmap).
- Do you judge prompt quality?
- No. Prompt debugging and eval playgrounds are someone else's territory. We track spend; we don't judge outputs.
- Which models are priced?
- The price book covers the major OpenAI/Anthropic/Google models; unknown models fall back to a default rate — never silently zero.