A router that actually forwards
One OpenAI-compatible endpoint for all your apps. AgentLedger picks the model, forwards the call — streaming included — and writes the measurement while doing so. Not a meter next to a proxy: the same movement.
Showcase · AgentLedger
Measuring, routing and accounting are one mechanism.
AgentLedger sits between your apps and the models. It picks the cheapest suitable model per task, actually forwards the call, and writes down what happened along the way: which route, which backend really executed it, how many tokens, what it cost — and when it does not know for sure. A meter next to a proxy is always one step behind the facts; this is the same movement.
Status: beta. Running in production in our own lab, not generally available yet. There is deliberately no release date — it follows once measurement coverage is complete.
67,309
calls in the ledger
8
model providers behind one router
9
apps recognised automatically
14
screens in the Mac app
Architecture
One passage, four things at once
The app talks to a single endpoint. Behind it the source is recognised, that source's policy applied, the call forwarded to the chosen backend and the result written down — including what the backend itself chose, because that is rarely the same as what was asked for.
What it does
Six things happening at the core
Routes per app
Set per app: automatic, a fixed model, let AgentLedger choose, or off entirely. An empty policy means exactly the old behaviour — you can switch it on without breaking anything.
A decision trail per call
For every call it is recorded why it went the way it did: which task type was recognised, which route setting applied, where it went, which backend actually executed it, and what that cost.
Honest pricing
Every call carries a pricing status: priced, free, subscription, local or unpriced. Unpriced is not a zero but a gap — and any total containing one is shown as a lower bound.
A spending brake
An hourly limit on real spend. Above it, paid calls go to the free model instead of running on. The brake fires on euros, never on subscription value.
Loop and ghost detection
Near-identical calls repeating inside a time window are flagged even when they are free — a free loop costs no money but does cost quota and time. Calls that hang are counted separately.
From the running app
Eight screens, unretouched
These are screenshots from the native Mac app in English, taken on 19 August 2026 with real production data from our own lab. Nothing was drawn in and nothing polished away — the failed calls and the unpriced gaps are simply there, because that is exactly what the product is about.
The core of the design
A gap must not become a zero
Almost every cost dashboard has the same silent fault: whatever it cannot price, it books as zero. The total then looks healthy and nobody sees that a piece is missing. AgentLedger does it the other way round.
pricedrate known, amount computed from the actual upstream
freeexplicitly € 0.00 in the price table — proven, not assumed
subscriptioncovered by the monthly fee; never triggers the spending brake
localruns on your own hardware; no invoice, but time and power
unpricedempty, not zero — and any total containing one is called a lower bound
What we got wrong along the way
Four measurement faults, and what they taught us
Each of these sat unnoticed for months inside a system that looked healthy. They are here because the correction is the product: these are exactly the places where a cost dashboard lies to you without knowing it.
A gap in the measurement was booked as € 0.00
Every call without a known rate silently got zero euros. On a single day, 14,911 calls sat at € 0.00 against 1.17 billion input tokens. The total was therefore not a cost figure but a lower bound, without anyone being able to see it. Since then “unpriced” is its own status with an empty value instead of a zero.
Source detection that had never once fired
Of 39,542 logged calls, all 39,542 sat under the source “http”. Every app was one app as far as the router was concerned, so per-app routes had nothing to grip on. Detection now runs in a fixed order — explicit header, own rule, sniffing — and records per call which rule actually fired.
Orphaned calls came from lock contention, not from crashes
Under heavy concurrent streaming the finalising write failed with “database is locked”, leaving a call stuck in status “routing”. Eighty-three of them, two to fifteen days old. The fix was in the database settings, not the logic — plus a safety net that closes any dangling row anyway. Nothing was deleted; they are marked as aborted.
Pricing the alias is guessing
If the router itself does not report which endpoint it became, the alias rate is an assumption. That assumption either costs tens of euros a day that do not exist, or it hides paid traffic. So pricing follows what the backend actually executed — or nothing at all.
From the ledger
Where the measurement stands today
67,309calls recorded, 63,099 of them successful
4,074failed calls kept with their reason — previously visible only in the error response
7.23 bninput tokens against 85.8 million output tokens (roughly 84 : 1)
9apps the router tells apart, each with its own recognition rule
€20.18measured cost across the 29 sessions in view, against €1,376.53 saved by routing
644 / 5,000calls without a rate in the session overview — made visible, not rounded away
The same correlation thread runs through our workflow automation.
Built with Python (stdlib)SQLite (WAL)SwiftUIMCPCloudflare Accesslaunchd
Beta — no release date yet
Want to know what your agent team really costs?
AgentLedger runs in production in our own lab and is in beta. There is deliberately no release date: measurement coverage per provider has to be finished first, because a measuring instrument that papers over its gaps is worse than none at all. Looking along or thinking along is welcome.