C Celladore
Celladore Docket · public alpha

Know where the cost actually fell.

Celladore Docket preserves provider-native billing evidence and is being verified to reconcile it against invoice-confirmed actuals, then attribute the result to the workload, customer, agent, feature, and outcome that caused it.

Sign in to Celladore Docket View on GitHub Read the docs
01 · ObserveProvider quantities and unitsCredits, allowances, exclusions
02 · ReconcileEstimate → observed → invoiceCorrections remain reproducible
03 · AttributeRoute and policy at event timeCoverage and confidence stay visible
04 · AllocateWorkload, customer, outcomeShared, idle, and residual balance

Actuals first, explanations attached

Estimates are useful, but they never overwrite observed quantities or invoiced amounts. Docket keeps each state and the evidence that moved it forward.

Tier 1 · ~1 min

Realtime metrics

Azure Monitor metrics surfaced directly in Grafana — no extra infrastructure required.

Tier 2 · 1–5 min

Near-realtime AI usage

Azure OpenAI diagnostic logs stream through Event Hub into a cost estimator, priced per request as it happens — aoai_cost_estimate() running live in ADX.

Tier 3 · 4–24h

Daily billed costs

FOCUS-format Cost Management exports land daily via a Container Apps Job at 3 AM UTC — the numbers that reconcile against the actual invoice.

AI Advisor · Weekly

Optimization recommendations

gpt-4.1-mini reviews all accumulated ADX cost data on a weekly cadence and produces ranked, actionable suggestions stored in cost_recommendations.

Cost Centre API

Adoption & lifecycle

FastAPI service tracking which teams and cost centres are onboarded. ADX-backed in production, Table Storage fallback for local dev.

PowerShell resource management suite

Eight production-grade scripts for the full resource lifecycle — capabilities vary by script, with advanced features like IaC detection and cascade depth limits available where appropriate.

Script Purpose
Invoke-ResourceInventoryList resources with cost and Azure Advisor enrichment
Invoke-ResourceDeletionSafe deletion with lock checks, IaC detection, and cascade depth limits
Invoke-ResourceLifecycleStop / start / restart with state preservation
Invoke-ResourceSuspensionSoft-disable: locks, network deny, scale-to-zero
Invoke-ResourceScalingSKU up/downscale with profiles and cost preview
Invoke-ResourceTaggingBulk tag apply, remove, audit, and sync from cost centres
Invoke-ResourceHealthAuditCompliance checks: tags, locks, orphans, naming conventions
Invoke-ResourceOrchestratorWorkflow composition: undo, history, cleanup

Built on

Provisioned end-to-end with Terraform; no manually-created resources.

# ingestion & storage
Azure Data Explorer (pvc-docket-kusto) · Event Hub · Storage Account (Table fallback)
 
# services
FastAPI (Cost Centre API) · Container Apps Jobs (daily export) · Grafana (5 dashboards)
 
# AI advisor
gpt-4.1-mini · ADX cost_recommendations table · weekly cadence
 
# infra
Terraform · GitHub Actions (OIDC) · Azure Container Apps · PowerShell 7.0+

Where it sits in the ecosystem

Celladore Docket is the evidence layer for operational economics across the portfolio. It consumes source facts without replacing their native units or provenance.

Sluice Primary data source Sluice emits AOAI diagnostic events that stream through Event Hub into Docket's ADX tables — priced per-request in near-realtime.
Retort Run attribution Correlates Retort agent run IDs with exact dollar amounts and token metrics for complete auditability per agent and per repo.
CodeFlow Per-PR cost tracking AutoPR AI spend tracked per PR analysis run — so you know the exact cost of each automated review cycle.
Baton Budget guardrails Docket budget thresholds surface in Baton's attention queue — blocking agent task dispatch before spend caps are breached.

Why it was built

Without agent-level cost attribution, runaway recursive agent loops across Mystira's story-generation pipeline were only discovered weeks later on aggregate monthly cloud invoices. There was no way to trace which agent, task, or repo caused the spike — just a number on a bill.

Docket was built to make LLM spend a first-class observable: per-agent, per-project, per-request — in near-realtime. The three-tier architecture exists because no single latency satisfies all three questions: "what's happening right now," "what did we spend this hour," and "what does the invoice actually say."

What it would cost

Docket has never been offered to anyone at a price. These are proposed tiers, not a bill — scoped to the cost centre, since that's the unit its own Cost Centre API already tracks, with a tracked-spend ceiling per tier so the plan also scales with what Docket is actually watching.

Starter
$49/mo

Up to 3 cost centres, up to $7,500/mo tracked spend

  • All three latency tiers, realtime through invoice-reconciled
  • 5 Grafana dashboards
  • Email support
Team
$149/mo

Up to 10 cost centres, up to $20,000/mo tracked spend

  • Everything in Starter
  • Weekly AI Advisor optimization recommendations
  • PowerShell resource-management suite
Fleet
Custom

Unlimited cost centres, unlimited tracked spend

  • Everything in Team
  • Cross-repo attribution — Sluice, Retort, CodeFlow, Baton
  • Dedicated onboarding

Would this get budget approved on your side? Tell us — tell us which tier (or cost centre count) you're interested in ↗ — spend ceilings are grounded in published cost-visibility-tool pricing, not invented.

Celladore Docket is preparing its public alpha. CLI, browser, and JSON contracts are available for verification. Access remains controlled while Mystira login and representative cost evidence are verified with legitimate sessions. Building similar FinOps tooling for agent fleets? Talk to us →