Steward
Sign in

Documentation › Sponsor guides

Tiers + credits

Steward sells four tiers. This doc explains what each one includes, how the credit budget that gates Deep Analysis actually works, and what happens when a tenant exceeds its daily cap. The pricing page has the marketing pitch; this is the operational reference for the sponsor (the buyer).

The four tiers

Tier Price Repos Seats Headline capability
Community $0 forever 1 public repo 1 Contributor-reputation scoring on PR open. Low-priority queue. Caliper under Business Source License (BSL) non-production terms.
Pro $19 / month 3 1 Full 9-dimension governance + Bring-Your-Own-Key (BYOK) toggle. High-priority queue. Caliper commercial-use license.
Team $79 / month 10 5 Shared workspace, basic audit log. Caliper commercial-use license.
Business $299 / month 25 25 Single Sign-On via OpenID Connect (SSO/OIDC) — Okta, Azure AD, Keycloak, Google Workspace. Audit log + Service Organization Control 2 (SOC2)-track reporting. Dedicated worker.

Annual billing gives you two months free across every paid tier. The Caliper commercial-use license is bundled in every paid tier and does not unlock additional surfaces beyond what its standalone plan offers — it's there so a Pro / Team / Business customer isn't paying for it separately.

What changes when you upgrade

  • Repo cap is the most-felt difference. A tenant on Pro with 3 covered repos who installs the App on a 4th sees the 4th land in "over cap" state — the App still receives webhooks but evaluations don't enqueue until you either uninstall one of the original 3 or upgrade.
  • Daily credit cap scales with the tier (see below). This is the operational guardrail; you'll hit it before the repo cap if your fleet is busy.
  • Seats matter for Team / Business — a seat is a member with org#read or higher on at least one of the tenant's covered orgs (see Org membership when that lands).
  • Audit + SSO unlock on Business and don't backfill — events before the upgrade aren't suddenly visible in the audit log.

Credits

A credit is Steward's unit of Large Language Model (LLM) spend accounting. One credit is approximately $0.05 of LLM cost. The cost per evaluation depends on the scan kind — see Credit usage for the three-scan model in detail. As a thumbnail:

  • A per-PR scoring call (the most common kind) is a flat 0.05 credits.
  • A Standard Scan (deterministic, repo-wide) is 0.5 credits before the language multiplier.
  • A Deep Analysis (full 9-dimension repo-wide eval) is 45 credits for an Ultra-Light language, 112.5 for Memory-Heavy, 135 for Compiler-Heavy.

Credits do not roll over across days. The counter resets at UTC midnight.

Per-tier daily caps

Tier Daily cap (credits) Approximate $/day Approximate per-PR scorings/day
Community 4 $0.20 ~80
Pro 40 $2.00 ~800
Team 160 $8.00 ~3,200
Business 600 $30.00 ~12,000

These are safety brakes, not subscription minutes. A normal month's evaluation traffic averages well below them. The cap exists so that a single bad day — runaway CI loop, mis- configured webhook firing on every push, a contributor opening 500 PRs in an hour — doesn't burn a year's worth of LLM budget in 90 minutes.

What happens when the cap hits

The tenant flips into Static-Only mode until UTC midnight. Concretely:

  • Standard Scans still run. The deterministic gates — Developer Certificate of Origin (DCO), Contributor License Agreement (CLA), complexity, Static Analysis Results Interchange Format (SARIF) ingestion, the path-based scorecard — keep firing exactly as before. They cost zero credits.
  • Deep Analysis blocks. Per-PR LLM-grounded reports (synth.GeneratePR) don't run. The system falls back to the deterministic-only report (synth.PRDryReport), which is the same thing it ships when the LLM provider has an outage. Your PR comment still appears; the prose is template-based rather than LLM-written.
  • Per-PR rows land as skipped(budget_exhausted) rather than failed in the dashboard. Distinct on purpose: a wave of skipped reads as "tenant hit the cap"; a wave of failed reads as "something broke." See Debug a failed evaluation for the distinction.
  • Manual evaluations triggered from /app/repos/... return a BUDGET_EXCEEDED envelope. The dashboard surfaces this clearly.

You don't get charged for the skipped calls — no LLM call is made, no credits spent. The cap is a circuit breaker, not a bill.

The cluster-wide brake

In addition to per-tenant caps, Steward enforces a cluster-wide $100/day = 2,000 credits disaster brake. If the combined spend across every tenant exceeds it, every tenant enters Static-Only mode, the on-call rotation pages, and we investigate. This sits well above normal cluster spend; you will never hit it under expected use. If you do, it's a Steward incident, not yours.

Choosing a tier

A quick decision tree:

  • Solo Open Source Software (OSS) maintainer or single-repo developer (public)? → Community. Costs nothing; you get reputation scoring and Standard Scans. The 4-credits/day cap covers ~80 Deep Analyses, enough for a typical small repo's weekly traffic.
  • Small portfolio (2–3 active repos)? → Pro. You want the 9-dimension governance + Bring Your Own Key (BYOK) to bring your own LLM provider key and pay-direct.
  • Startup with private + public repos and no OSS Program Manager? → Team. The 5-seat shared workspace is the load-bearing capability; basic audit log covers the compliance basics.
  • Commercial Open Source Software (COSS) company, regulated org, or anything where compliance is an explicit requirement? → Business. SSO + SOC2-track reporting are gated to this tier; the dedicated worker means your per-tenant queue doesn't share load with anyone else's.

Bring-Your-Own-Key (BYOK)

Pro and above can toggle BYOK to point the per-PR LLM scorer at your own Anthropic / OpenAI / etc. account. When BYOK is on:

  • Credits don't accrue. Every LLM call goes through your provider's billing; Steward bills you only for the subscription itself.
  • The daily cap still exists — but it counts your provider's calls so you can see your own burn from /sponsor/usage. The cap as a circuit breaker stays useful; if you really want it off, talk to staff.
  • Your provider's rate limits become the gating factor. If the provider 429s, the per-PR worker falls back to the deterministic report just like for a regular outage.

BYOK is the right setup for a tenant whose total LLM spend will exceed their plan's bundled credit cap by enough that paying the LLM provider directly is materially cheaper than letting Steward mark it up.

What's next