Everything people ask before signing up.
Pricing, security, integrations, free tools, refunds, support — all answered straight, no marketing fluff. If your question isn't here, email hello@cartieai.com and we'll add it.
A Financial Operating System for AI and Cloud spend. We start with a drop-in proxy that cuts your OpenAI / Anthropic / Gemini / Groq / Mistral token bill by 20-50% (up to 70% on repetitive workloads), then 20 more engines hunt waste across AWS, Azure, GCP, Snowflake, and Databricks — and ship the fix as a GitHub PR.
Engineering leads, founders, and CFOs whose AI token bill suddenly got bigger than their EC2 bill. Sweet spot: teams running $5K–$500K/mo on LLM APIs and $10K–$5M/mo across AWS / Azure / GCP / Snowflake / Databricks who want one place to see, explain, and shrink every dollar.
Five AI-economics engines (Reasoning Efficiency, Prompt Compactor, Semantic Cache + Miss Hunter, RAG Embedding Cache, Real-Time Token Arbitrage), three world-first engines (AI Agent Economics, Embedding Store Audit, Pipeline DAG Dedup), five cloud-cost engines (Bill Detective, Idle Resource Killer, K8s Shapley Attribution, Egress Predictor, Auto-Tagger), and the action layer (PR Cost Predictor, Decision Simulator, Anomaly Detection v2, Cost Bisect, Carbon-Cost Pareto, Shadow AI Audit, CARTIE Lens, Budget Negotiator™). The same list powers our public marquee and the /honest-scorecard.
When your spend spikes, we attribute every extra dollar to a service-level driver (rate, volume, region, time-of-day) using deterministic Python — no LLM hallucination. Claude Haiku writes the 3-bullet explanation on top of those numbers. We show our work: every claim links to the underlying billing line items.
It can't — we never write to your infra. We open a GitHub Pull Request against your IaC repo with the Terraform/CDK diff. Your existing review process, CI, and approvers apply. We default to read-only IAM. The PR includes a savings estimate + a rollback plan. You merge or close, like any other PR.
Yes — it's a deterministic PPTX/PDF generator (not a 'here's a chart' export). 12 slides: monthly spend, runway, top-3 anomalies with root cause, vendor renewals coming up with our negotiated benchmarks, AI feature P&L, and 90-day forecast. CFOs hand it to the board with their own commentary on the title slide.
Three signals must agree: (1) spend-per-user diverges from baseline by ≥3 stddev, (2) prompt entropy collapses (same prompt firing repeatedly = retry storm or bug), (3) cache-hit ratio drops below user norm. Each is independently observable — false positives are rare and always actionable.
Plain-English cost questions: `/cartie why is aws up 14%?` returns a 3-bullet root cause in under 2 seconds. `/cartie top 3 recommendations`, `/cartie burning` (live $/hr), `/cartie envelopes` (per-team budget). Free-form questions route to Claude Sonnet. Audit-logged, signing-secret verified, SOC 2-aligned.
Yes — we ship an MCP (Model Context Protocol) server inside our npm CLI. After `npm install -g @cartieai/cli && cartie init`, the bot is available as a tool inside any MCP-compatible client. Tools: `cartie_why`, `cartie_visits`, `cartie_engines`. Your AI assistant can answer cost questions without leaving the editor.
Two ways. (1) Live token-savings demo at cartieai.com/try — paste your own prompt, see the exact tokens + dollars we'd save, no signup. (2) Click "Try live demo (no signup)" on the homepage and we drop you into a fully populated demo dashboard in one second with realistic seed data. No credit card, no email harvesting.
Two timelines. **Time-to-data:** under 10 minutes for the AI feature side (POST events from your code with an API key) and 24 hours for the full cloud-cost side (upload a CUR/Azure/GCP export — we don't auto-pull because most enterprise teams prefer upload-only for security). **Time-to-savings:** typically 2–3 weeks before the first verified savings event hits the dashboard.
Yes — AWS, Azure, GCP, Snowflake, Databricks, and DigitalOcean. Bills are normalised into the FOCUS open-standard so a 'compute hour' means the same thing across providers. Cross-cloud egress (e.g., GCP Snowflake → AWS Bedrock) gets its own predictor.
Cost-tracking covers OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Google Vertex AI, Cohere, Mistral, Groq, Together, and DeepSeek. The drop-in Auto-Compiler Proxy specifically supports OpenAI, Anthropic, Google Gemini, Groq, and Mistral today — all OpenAI-compatible, so the same client code works for all five. Auto-detects provider/model/feature spans without code changes.
Yes — Kubernetes via cAdvisor scrape, Snowflake via its `account_usage.warehouse_metering_history` view, Databricks via the system tables export. All three feed the same Customer P&L so you can attribute spend per feature, per customer, per team — across the entire data stack.
The web app is fully responsive and we ship a daily Slack 1-liner so you don't need to open a dashboard. A native PWA install prompt is also available if you want CARTIE on your home screen.
Two tiers. **Pilot Tier ($497/mo, 100 seats)** — locked for 12 months, month-to-month cancel, full product, no feature gates. **Standard Tier ($1,997/mo)** — same product, no seat cap. Both are flat fees, not % of savings. We'd rather build a long relationship than charge a percentage of someone else's success.
If we don't save you at least 3× your annual subscription within 90 days of go-live, we refund every dollar you paid. "Save" = verified savings logged in your dashboard, signed off by your team lead. It's a clause in the master services agreement, not a marketing line. Refund processes within 14 days of the 90-day mark.
Full refund. The pledge is binary — 3× or money back, no pro-rating. Yes, this means we sometimes refund deals we technically saved money on. We accept that because it forces our sales team to qualify hard, and our engineering team to ship engines that actually deliver. 2× isn't worth the friction.
Your **price** is locked at $497/mo for 12 months — we can't raise it. Your **commitment** is month-to-month: cancel anytime, no termination fee, export your data, rotate keys, done. "Locked" only protects you, never us.
All public calculators (savings, bill score, AI cost calculator, commitment coach, PR cost predictor, decision simulator, snowflake audit) are 100% free with no signup. The full multi-tenant product unlocks at the $497/mo Pilot Tier.
Stripe metered billing on a 30-day-net invoice. Every line item is auditable — no surprise charges, no opaque platform fees. We use Stripe Test Mode during the pilot phase; production switches to live keys at GA.
Yes — month-to-month, no contracts, no termination fees. Export your data on the way out (CSV / JSON / FOCUS exports) and rotate your Stripe restricted key to lock us out instantly.
Yes — within 14 days of any invoice if you can show a billing error or duplicate charge. The 3× ROI pledge is a separate, broader refund mechanism that triggers at 90 days. Detailed refund policy on /refund.
Those tools either auto-manage commitments (ProsperOps) or surface anomalies and dashboards (Cloudability/Apptio). CARTIE answers the questions they don't: which AI features pay for themselves, which customers cost more than they pay you, what the next $1 of cost will buy, and ships the fix as a GitHub PR. Decision-support + action, not just dashboards.
Vantage is a cost visualizer — beautiful charts, no AI cost coverage, no action layer. CARTIE adds the missing 60%: the AI/LLM stack (tokens, prompts, cache, eval), the GitHub-PR action layer, and the CFO-facing P&L. Most teams keep Vantage for cloud cost viz and add CARTIE for AI cost + execution.
Finout tracks AI cost as a category. CARTIE goes deeper: per-feature unit economics, per-customer P&L, token leak detection, prompt-level cost attribution, semantic caching that actually returns $0 on cache hits, and the action layer (PR-time guardrails, one-click fixes). Plus the 3× ROI refund pledge — Finout doesn't have one.
Yes for the cost-attribution side — we read the same cAdvisor + kube-state-metrics signals KubeCost does. We're not a 1:1 replacement on the Kubernetes-specific deep features (node autoscaler advice, custom-resource allocation), but most teams find that 80% of KubeCost's value is the per-namespace cost view, which CARTIE matches at $497/mo vs KubeCost's $1,000+/mo enterprise tier.
Helicone is a great proxy for trace logging — but it stops at "here's your token use." CARTIE picks up where Helicone ends: semantic caching that saves $0 on hits (Helicone doesn't), prompt compaction that cuts 20-50% of tokens on typical prompts (up to 70% on repetitive workloads), per-feature/customer attribution, eval-driven model swaps, and the Slack/CLI/MCP surfaces. Our scorecard: 20/21 vs Helicone 1/21 on these.
Three honest cases: (1) Your cloud spend is under $10K/mo — the ROI math doesn't work, use our free tools instead. (2) You don't have an AI/LLM workload yet — half our engines won't fire and you'd be paying for capacity you can't use. (3) Your team won't merge PRs we write — we're an action layer, and if your culture is read-only-dashboards, the value drops 60%.
Under one engineer-week for the full migration. Day 1: connect cloud (CUR upload + Stripe read-only key + Slack OAuth — 30 min). Days 2–3: install the LLM Proxy or POST AI cost events (5-line code change). Day 4: set policies + invite team. Day 5: cancel your previous tool. We migrate Vantage/Finout/Apptio dashboards on request — share an export and we recreate them.
Not yet — and we'll never claim otherwise. Today we operate with SOC 2-aligned controls (15 written policies derived from the open-source strongdm/comply MIT-licensed bundle) and our auditor selection begins at customer #25. We publish the policy bundle and live security proof for your InfoSec review.
Yes for the rights customers can exercise (data access, export, deletion, portability) — those work end-to-end today. EU data-residency and a formal DPA are available for Enterprise customers; reach out to hello@cartieai.com.
30 days for raw cost-line events by default, 90 days for aggregated metrics, indefinitely for your savings and verified-action history (you control deletion). Audit logs are retained 90 days minimum to satisfy compliance.
AES-256 at rest in MongoDB Atlas, TLS 1.2+ in transit, Cloudflare-fronted ingress. Stripe API keys are stored encrypted with envelope encryption — even our database backups can't read them without a key the application holds in memory.
Stripe (payments), Resend (transactional email), Slack (alerts), MongoDB Atlas (database), Cloudflare (CDN/WAF), and our LLM providers (OpenAI / Anthropic / Google) for narrative generation only. Full list with purpose & data-shared is gated to verified InfoSec reviewers.
US-East (Virginia) by default. EU-West and APAC residency are available for Enterprise tiers — your tenant rows live in a geographic shard, our application enforces it at the DB-collection level.
Yes — admin role can pull JSON or CSV exports from /admin/audit-log with date filtering, plus on-demand security-proof PDFs that show the live tenant-isolation checks running on your account.
Internal red-team review is run weekly by our automated leak-guard cron (4 rounds + 1 weekly). External 3rd-party pen test is scheduled alongside the SOC 2 Type II audit at customer #25 — we publish the dated commitment on /trust.
The demo tenant is read-only and runs in a sandboxed shard with no real customer data. Every demo session shares the same seed data and resets on logout. There's nothing to leak. Production tenants run on isolated shards with strict cross-tenant guards (live proof on /admin/security-proof).
No. We never share your data with model providers in a way that could be used to train models on it for other customers. LLM calls (Claude / GPT / Gemini) only ever see anonymised, structured numbers — not customer names, emails, or workload identifiers.
Only restricted keys with read-only Customer/Subscription/Invoice scopes (90-day window). If you'd rather not share a key, the CSV-upload path imports the same data without one. Either way, the key is encrypted at rest, never logged, and revocable instantly from your Stripe dashboard.
Email addresses (yours and your customers' — for P&L attribution), names if you provide them, and IP-address-derived org metadata. No payment-card data, no source code, no production database content.
Yes — every list and table has a CSV export, and admin-tier users can pull a full JSON dump of all tenant collections. Deletion is one click from /settings — verifiably wipes your tenant rows within 24 hours.
Physically scoped at the database level — every read/write is proxied through tenant_db.py which forces the tenant filter at the collection-name level, not in-memory. We publish a live test that fails if a tenant ever sees another tenant's data.
Generate an API key on /settings/api-keys, set it as CARTIE_API_KEY, and POST to /api/ai-cost-events/track with {feature_slug, cost_usd, customer_id}. Python, JavaScript, and cURL snippets are on /developers.
Yes — 1000 events/minute per API key on the single-event endpoint, 50 batches/minute on /track-batch (each batch up to 500 events). 429 with retry-after header on overflow. Higher limits available on Enterprise.
Every outbound webhook carries an X-Cartie-Signature: sha256=<hex> header. Recompute hmac.sha256(signing_secret, '<timestamp>.<raw_body>') and timing-safe compare. The signing secret is shown once when you create an API key.
Stripe (revenue sync), Slack (alerts/digests/bot), Resend (email), AWS CUR / Azure / GCP / Snowflake / Databricks / DigitalOcean (cost imports), GitHub (Auto-PR + Actions for PR-time cost), Claude Desktop / Cursor / Continue (MCP server). VS Code extension is in beta.
Yes — Google SSO is live for all tiers; Okta / Azure AD / SAML / SCIM provisioning are available on Enterprise. Reach out to hello@cartieai.com to get configured.
Genuinely free, no signup, no rate limit. Some have an optional 'send me the detailed plan' email field — if you submit it, we add you to the appropriate sales lane (CFO / engineer / InfoSec). That's the only catch. Don't submit the email and we never see who you are.
It uses the same 18% FinOps Foundation baseline + driver multipliers most enterprise audits use. Capped at 40% so we never over-promise. Real audits land within ±5pp for ~80% of teams.
Yes — paste any `terraform show -json` output, see the exact dollar delta in 2 seconds. Supports EC2 (29 instance types), RDS, ElastiCache, Lambda, NAT GW, EBS, S3 — priced from a Q2-2026 us-east-1 snapshot. Architecture-aware (gp3 vs gp2, Graviton vs x86).
Yes — every result generates a sharable badge SVG with your score. Tweet it, drop it in Slack, or paste it in your engineering README. The badge links back to the quiz so colleagues can compare.
100% deterministic Python. Every scenario (LLM swap, SP coverage increase, caching, K8s rightsizing, drop unprofitable customer, token cap) is a pure function with side-effect-free math. The LLM only writes the plain-language narrative on top of the numbers — never the numbers themselves.
Email hello@cartieai.com — replies within 4 business hours, usually faster. Pilot/Standard customers get a private Slack Connect channel and a named contact. We don't gate basic support behind a paid tier.
Yes — first login fires a 7-step product tour (Anomaly Autopsy · One-Click Fix · AI CFO · Wall of Tenants · Slack bot · CLI/MCP · 26-engine catalog). You can replay it from /settings any time. The Daily AI Brief widget surfaces the most useful next-step on each visit.
Yes — Pilot Tier comes with 100 seats, Standard Tier is unlimited. Roles are admin / member / viewer. Engineers can log savings actions on the Engineer Savings Wall without admin permissions, and the verify step keeps numbers honest.
Best-effort 99.9% uptime today (live status on /system-status). A formal contractual SLA with credits is available on Enterprise tier. We don't fake an SLA we can't deliver.
Email hello@cartieai.com or shout it on the founder's calendar (linked in /contact). We ship roadmap-public — every accepted request gets a public ETA and a Slack ping when it ships.
Founded by Lakshmi Kiranmai Guduru. We build in public — see /about for the story and /careers if you want to join.
Bootstrapped, Texas-registered LLC, debt-free, founder-funded through customer #50. We're not chasing the next round before we have $1M ARR — that's how 80% of YC FinOps startups died. If we shut down, source code is committed to escrow and your data exports stay live for 12 months.
Q3 2026: GA on SOC 2 Type I, VS Code extension out of beta, OpenTelemetry-native span ingestion. Q4 2026: SOC 2 Type II audit kickoff, Kubecost feature parity, multi-currency budgeting in Customer P&L. We publish every accepted feature with a public ETA — request yours via /contact.
It's the financial OS for the cloud — every line item, every fix, every refund tracked on one cart. Plus the .ai domain was available and our auditor couldn't pronounce "Vantage".
It's our live 60-second token-savings demo. Paste your own prompt, pick a model (OpenAI / Anthropic / Gemini / Groq / Mistral), and our real production compactor runs the exact same six passes our proxy uses — then shows you the exact tokens saved, dollars saved, and monthly savings at your call volume. No signup, no card, no email. The compactor processing the demo prompt is the same code path that processes paying customers' traffic.
It's the ceiling, not the average. The Auto-Compiler Proxy uses 6 deterministic passes: whitespace normalization, comment stripping, boilerplate pruning, sentence dedup, long-block truncation (only above 3K tokens), and content normalization. On clean prompts we typically save 5-20%. On average system-message-heavy prompts: 20-40%. On repetitive or boilerplate-heavy prompts (agent loops, RAG retries): 40-70%. Layer in semantic caching + model cascade and the whole-system savings ceiling is 60-70% — which is the figure you'll see on the homepage. Try it on your own prompt at /try and you'll see your actual number.
Five today: OpenAI, Anthropic, Google Gemini, Groq, and Mistral — all OpenAI-compatible at the API level so the same client code works for every provider. Just point `base_url` at https://api.cartieai.com/v1 and add your CARTIE proxy key. Cohere, AWS Bedrock, and Azure OpenAI are cost-tracked today (via SDK telemetry) and proxy support is on the Q3 2026 roadmap.
Not today — we ship as a managed SaaS to keep velocity high and security tight. An on-prem / VPC-deployed option (Helm chart + air-gapped container registry) is on the roadmap for Enterprise customers at $50K+/year. Until then, our zero-data-retention defaults + read-only IAM + EU residency cover the same threat model for 95% of teams.
HIPAA-eligible (BAA available on Enterprise after SOC 2 Type II — Q1 2027). FedRAMP and FINMA require a sovereign-cloud deployment which we don't offer today — talk to us if you're an early customer who needs it. We won't fake a badge we haven't earned.
All LLM-powered features (Anomaly Autopsy, AI CFO, "Why?" explainer, Slack bot) fall back to deterministic canned responses + a banner indicating degraded mode. Core engines (Prompt Compactor, Idle Resource Killer, Bill Detective, PR Cost Predictor, Cost Bisect) are pure Python — they don't use LLMs at all and stay 100% functional during outages. The Auto-Compiler Proxy itself also stays online — it just falls back to your chosen upstream if one provider is down, and the compaction/cache layers keep saving regardless.
All four. One-Click Fix emits the dialect matching your repo: HCL for Terraform, TypeScript/Python for CDK, Pulumi's native SDK for Pulumi, and YAML for Crossplane. We auto-detect from the repo structure on first PR — no config needed.
Yes — Enterprise tier supports BYO LLM keys for all narrative-generation features (Anomaly Autopsy, AI CFO, Slack bot, "Why?"). The keys are stored encrypted with envelope encryption, never logged, and used only for your tenant's calls. Default tier uses our shared Universal LLM key billed to us.
We treat every user-supplied message as untrusted input. The bot routes pleasantries + canned queries through deterministic intent detection (no LLM exposure); only free-form questions hit the LLM, with strict system-prompt boundaries and output filters that block any "ignore previous instructions" patterns. We log + alert on any suspected injection attempt.
We log the invocation (user, channel, timestamp, intent) for audit purposes, but not the full prompt text or the LLM response unless you opt in to debug logging. Even with debug on, payloads are auto-redacted of emails/UUIDs/credentials before storage. Logs are retained 90 days max.
Yes — Enterprise tier supports EUR, GBP, JPY, INR, AUD, CAD, CHF, SGD with daily exchange-rate sync from the ECB feed. All dashboards convert to your org's primary currency; multi-currency budgets are a Q4 2026 feature.
Yes — every account gets a separate `staging` tenant alongside production. Use it to validate PR-time guardrails, eval new engines, or test webhook handlers without touching real data. Sandbox tenants don't count toward seat limits.
Yes — Partner Tier ($4,997/mo) lets MSPs and consultants run a multi-tenant CARTIE instance under their own brand. You bring the customers, we provide the engine. Reach out to hello@cartieai.com — we have 3 partner slots reserved for design partners through end of 2026.
Daily MongoDB Atlas snapshots with 30-day retention. RPO 1h, RTO 4h on the primary region. Failover to us-west-2 is one-command DNS swap; we test it monthly. Source code escrow + 12-month data retention if we shut down the company.
Yes — every CFO-facing surface (AI CFO Advisor, Board Deck Generator, Customer P&L) is written in plain English with hover-tooltips explaining every term. The Board Deck Generator specifically converts FinOps metrics into the language finance teams already use: gross margin, contribution margin, runway impact.
Three things make the difference: (1) Slack `@cartie` bot — engineers don't have to leave the channel; (2) CLI/MCP server — works inside Cursor and Claude Desktop where they already live; (3) One-Click Fix opens GitHub PRs in their normal review flow. We don't ask them to learn a new dashboard. Adoption averages 84% in pilot teams within 30 days.
Yes for Pilot and Standard tiers — quarterly 60-min Zoom call with your named contact: savings recap, missed opportunities, next-90-day plan, and a written 1-pager you can share internally. Enterprise tier adds a monthly cadence.
Yes — 20% recurring commission for the first 12 months on any customer you refer that signs Standard or Enterprise. Sign up at /partners. We pay monthly via Stripe Payouts on a 30-day-net basis.
Today only English. Spanish and Portuguese (LATAM) are scheduled for Q3 2026, German and French for Q4 2026. Reach out if you need a specific language earlier — we'll prioritize based on customer demand.
No hard limit on Pilot (100 seats) or Standard (unlimited) tiers — every team member can be logged in simultaneously. We've stress-tested 500 concurrent active users in a single tenant with sub-200ms response times.