Back to Registry

Unified Cost Dashboard Plan (Docs-Verified Draft)

Objective

Build a single cost dashboard that tracks spend and usage across:

• LLM providers (OpenAI, Anthropic, OpenRouter, Google Gemini)
• Hosting/platform providers (Vercel, Cloudflare, WP Engine)
• Attribution targets down to app/site where possible

This plan is based on current provider docs/changelogs reviewed during this session.

---

1) Feasibility Matrix (by provider)

| Provider | Programmatic Cost/Usage Access | Granularity Confidence | Key/Auth Notes | Implementation Notes | |---|---|---|---|---| | OpenAI | Yes (Usage + Costs org endpoints shown in official cookbook examples) | High | Admin-level org key required for org usage/cost endpoints | Strong API path for time-bucketed costs and grouping | | Anthropic | Yes (Usage & Cost Admin API) | High | Admin API key (sk-ant-admin...) required; not available for individual accounts | Excellent for org-level reconciliation and monitoring | | OpenRouter | Partial-strong (generation-level usage/cost documented; account rollups less explicit) | Medium | Standard API key + generation IDs | Use per-request usage/cost and aggregate internally | | Google Gemini API / AI Studio | Billing docs exist; programmatic export path depends on project/billing setup | Medium-low | Depends on AI Studio/Cloud Billing setup | Likely requires Cloud Billing export integration for robust reporting | | Vercel | Yes (/billing/charges in changelog/API, FOCUS v1.3 stream) | High | Team token with billing access | Great source for FinOps-grade ingest | | Cloudflare | Usage/billing docs available; usage-heavy per-service metrics; invoice remains source of truth | Medium | Account token with relevant scopes | Good for usage; cost allocation may need internal mapping model | | WP Engine | API exists for platform/account ops; billing/overage data not clearly exposed as rich API in docs checked | Low-medium | Per-user API credentials | Likely combine API + portal export/manual feed initially |

---

2) Evidence Notes (docs checked)

• OpenAI usage/cost cookbook example includes v1/organization/usage/completions and v1/organization/costs patterns.
• Anthropic docs explicitly provide a Usage & Cost Admin API and require admin keys.
• Vercel changelog (Feb 19, 2026) confirms /billing/charges endpoint in FOCUS v1.3 format and CLI vercel usage.
• Cloudflare billing docs state usage-based billing and that invoice is most reliable source for billing info.
• WP Engine support docs confirm API focus is account/platform management and list many non-supported operations; billing API surface not clearly evidenced from reviewed pages.

---

3) Architecture Proposal

3.1 Data pipeline

1. Ingestors per provider (scheduled)
2. Normalize into one schema
3. Attribution mapper (team/app/site/env/project)
4. Cost calculator (when only usage is available)
5. Dashboard/API layer

3.2 Core normalized schema

• provider (openai, anthropic, openrouter, vercel, cloudflare, wpengine, google)
• account_id
• project_id / team_id
• site_id (nullable)
• service (model, workers, builds, etc.)
• metric_type (usage|cost|credit)
• quantity
• unit
• unit_price (nullable when direct cost exists)
• amount_usd
• period_start, period_end
• source_confidence (high/med/low)
• attribution_confidence (high/med/low)

3.3 Attribution model (critical)

• Prefer provider-native project/team tags first.
• Add internal mapping table:
  • provider account -> FuelVM app -> client website
• Keep manual override support for ambiguous spend.

---

4) Security and access model

Use dual-plane model from fuelvm-access-matrix-v1.md:

• Internal plane: full ingestion and admin controls
• External plane: read-only client-facing cost slices

Token handling:

• Store provider keys in secret manager only
• Rotate quarterly (or on exposure)
• Separate read-only billing keys from write/admin keys where possible

---

5) Phased implementation

Phase 0 (1-2 days) — Validation Spike

• Validate live pull for: OpenAI, Anthropic, Vercel, Cloudflare
• Confirm exact scopes and retention windows
• Output: feasibility report with real sample rows

Phase 1 (1 week) — MVP Dashboard

• Providers: OpenAI + Anthropic + Vercel + Cloudflare
• Features:
  • daily cost trend
  • provider breakdown
  • app/site attribution table
  • anomaly alert thresholding

Phase 2 (1 week) — Expand connectors

• Add OpenRouter and Google path (direct or modeled)
• Add WP Engine via best available source (API/export hybrid)

Phase 3 (1 week) — Client-grade views

• Per-client/site rollups
• Budget vs actual
• Forecasting (7/30-day burn)

---

6) Known unknowns / risks

1. OpenRouter account-level billing API depth vs per-generation reconstruction.
2. Google AI Studio vs Cloud Billing export path for robust automation.
3. WP Engine billing granularity via API appears limited from docs reviewed.
4. Cross-provider timestamp normalization and FX/tax treatment (if needed).

Mitigation:

• Mark each data source with confidence tags.
• Separate “invoice-accurate” vs “estimated” views.

---

7) Immediate next steps (actionable)

1. Create connector spike tickets for 4 high-confidence providers:
   • OpenAI, Anthropic, Vercel, Cloudflare
2. Define secret scopes per provider
3. Stand up normalized cost_events table
4. Build first dashboard page:
   • Total spend MTD
   • Spend by provider
   • Spend by app/site
   • Confidence badges

---

8) Recommendation

Build this now as a staged FinOps product:

• It is feasible.
• Highest confidence starts with OpenAI/Anthropic/Vercel.
• Cloudflare and OpenRouter are viable with careful attribution.
• WP Engine and some Google paths may start as hybrid (API + export/model).

This is not pie-in-the-sky, it is a practical 3-phase build with clear early value.