costQL

Price GraphQL queries before they run. costQL calibrates a live GraphQL API into a pricing pack, one self-contained JSON file, then quotes any query against that pack fully offline. No server, no network, no measurement.

pip install costql        # Python: build packs + quote
npm install costql        # JS/TS: quote packs (build stays in Python)

60 seconds to a quote (offline)

from costql import PricingPack

pack = PricingPack.demo("tmdb_t3")   # a demo pack bundled in the package
quote = pack.quote('{ movie(id:"27205"){ cast(limit:8){ person{ name } } } }')

quote["price"]        # safe billable ceiling, in cost-units (never dollars)
quote["typical_price"]# fair average estimate
quote["confidence"]   # high | medium | low: cyclic queries are flagged, not billed

Or from the command line:

costql quote --demo tmdb_t3 '{ movie(id:"27205"){ title } }'

(Quoting your own API? Build a pack with costql build and pass it with --pack your_pack.json. The bundled --demo pack is just for the tour.)

Every result follows a frozen output contract (v1.0): price is always present, always a number, always safe to bill on. See docs/contract.md.

How it works

1. Build (seller side, once per schema): run costql probe <url> to see what your endpoint supports, write a ~90-line adapter that tells costQL where your API is and how to fill in its arguments (or hand that job to a coding agent: see docs/agents.md), then run costql build. costQL introspects the schema, measures a set of clean calibration queries, and fits a per-resolver cost model.
2. Ship the pack: the output is one static file (schema + fitted costs + any observed outside hosts). Vendor it into any app.
3. Quote (app side, forever): load the pack, price queries locally; in Python or JavaScript.

One currency, three fidelities

costQL prices in work-ms, the summed duration of the real work a query causes. Three fidelities of one engine, set by how much your API's instrumentation exposes:

Honesty first: T1 works black-box against any GraphQL API: the Rick & Morty case study hit ~4% mean error with a 93-line adapter and zero server changes. T2/T3 require your server to emit costQL's cost-trace extension; the demo packs are mostly T3 because we instrumented the demo servers. Your first pack will be T1: that is the designed starting point, and it already gives you a safe billable ceiling.

Measured accuracy

On held-out queries against real backends (methodology in docs/evaluation.md: calibration and evaluation sets are disjoint; there is no query→price lookup):

• TMDB demo (instrumented): mean error T1 17% / T2 11% / T3 11%
• Rick & Morty (public API, not ours): ~4% mean error at T1, ceiling never under the real cost
• Northwind (batch-heavy SQLite): heavy entity sharing, the API shape that needs the sharing-watching tier: 12% mean error at T3 on hub queries that a sharing-blind fidelity cannot price (case study)

Cyclic-recursion queries (whose runtime dedup is unknowable pre-execution) are automatically flagged low confidence and priced as a structural ceiling. Flagged, not silently billed.

What costQL is not

• Not a service. The pack is static and local by design: no sidecar, no pricing endpoint, no extra API call.
• Not billing. costQL speaks cost-units, never dollars; converting to money is the consuming app's job.
• Not a rate limiter. It prices; what you do with the price is up to you.

Docs

Quickstart, the adapter guide, tier fidelity, the output contract, evaluation methodology, and an interactive playground: https://costql.com

License

Apache-2.0