> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tensorcost.com/llms.txt
> Use this file to discover all available pages before exploring further.

# TensorCost

> The control plane for every dollar your company spends on AI — across GPUs, managed inference, and agent workloads.

# TensorCost

**The control plane for every dollar your company spends on AI.**

TensorCost is AI cost governance across self-hosted GPU fleets, managed-inference APIs (Bedrock, Azure OpenAI, Vertex, OpenAI, Anthropic), and the agent workloads that drive both. One pane of glass. Real attribution. Recommendations that pay for the platform.

## Start here

<Columns cols={2}>
  <Card title="Get started" icon="rocket" href="/getting-started">
    Sign up, connect your first cloud account, see your first recommendation in 48 hours.
  </Card>

  <Card title="Install agents" icon="download" href="/agent-installation">
    A Docker container with console-issued credentials. IMDSv2 auto-detect, opt-in HMAC-authenticated streaming.
  </Card>

  <Card title="Connect Bedrock" icon="aws" href="/bedrock-integration">
    The lead managed-inference adapter. Routing, prompt cache, provisioned-throughput, runaway-loop alerts.
  </Card>

  <Card title="Connect cloud accounts" icon="cloud" href="/connect-cloud-accounts">
    AWS, Azure, and GCP infrastructure spend — separate from AI usage, read-only, verified before it activates.
  </Card>
</Columns>

## Coverage matrix

TensorCost is the only governance layer that spans all three sides of the modern AI bill.

| Workload class           | Examples                                                                              | What we attribute                                                 | How we ingest                                                                                                     |
| ------------------------ | ------------------------------------------------------------------------------------- | ----------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| **GPU fleets**           | A100, H100, H200, B200; MIG slices; NVLink topologies; on-prem Slurm/Ray; EKS/GKE/AKS | Per-instance, per-MIG-slice, per-namespace, per-team              | Unified GPU agent (gRPC + HMAC)                                                                                   |
| **Managed inference**    | Amazon Bedrock, Azure OpenAI, Vertex AI, OpenAI API, Anthropic API                    | Per-model, per-application, per-team, per-user                    | CUR 2.0, CloudWatch metrics, invocation logs (read-only IAM)                                                      |
| **Agent workloads**      | LangGraph, CrewAI, in-house orchestrators                                             | Per-agent, per-workflow, per-conversation; runaway-loop detection | Same data plane as managed inference                                                                              |
| **Cloud infrastructure** | AWS, Azure, GCP — compute, storage, networking, everything else on the bill           | Per-account, per-subscription, per-project                        | Cross-account IAM role (AWS), FOCUS Cost Management export (Azure), BigQuery billing export (GCP) — all read-only |

## Three SKUs

<Columns cols={3}>
  <Card title="GPU FinOps" icon="microchip">
    The original product. Fleet rightsizing, MIG slicing, spot blending, RI/Savings Plan optimization.
  </Card>

  <Card title="AI Inference Cost Governance" icon="brain">
    Bedrock + Azure OpenAI + Vertex + OpenAI + Anthropic, with model routing and prompt-cache recommenders.
  </Card>

  <Card title="Agent Cost Observability" icon="robot">
    Per-agent, per-workflow attribution. Loop and retry-storm anomaly detection. PagerDuty + Slack hooks.
  </Card>
</Columns>

## Explore

<Columns cols={2}>
  <Card title="Architecture" icon="sitemap" href="/architecture">
    15 backend services, 14 microfrontends, 21 packages. NestJS + gRPC + Postgres + Redis on Fargate.
  </Card>

  <Card title="API reference" icon="terminal" href="/api-reference">
    REST gateway, JWT-auth, tenant-scoped, versioned at `/v1/`.
  </Card>

  <Card title="CLI" icon="terminal" href="/cli">
    `@tensorcost/cli` — offline prompt-cost estimates and before/after comparisons from your terminal or CI.
  </Card>

  <Card title="Developer tools" icon="code" href="/developer-tools">
    `@tensorcost/sdk` (Node) and `tensorcost` (Python) — one-line wrappers for OpenAI, Anthropic, and Bedrock clients.
  </Card>

  <Card title="Real-time events" icon="bolt" href="/realtime-events">
    Tenant-scoped socket.io for live recommendations, alerts, and agent fleet state.
  </Card>

  <Card title="ML & anomaly detection" icon="brain" href="/ml-anomaly-detection">
    Burn-rate alerts, forecasts, and the four shipped Bedrock recommenders.
  </Card>

  <Card title="Observability" icon="chart-line" href="/observability">
    OpenTelemetry, structured logs, customer-side audit trails.
  </Card>

  <Card title="Configuration" icon="gear" href="/configuration">
    RBAC, tag mapping, budget hierarchies, alert routes, feature flags.
  </Card>

  <Card title="Feature flags" icon="toggle-on" href="/feature-flags">
    LaunchDarkly + `useFeature()`. Quarterly stale-flag cleanup ritual.
  </Card>

  <Card title="Compliance & SOC 2" icon="shield" href="/soc2-readiness-guide">
    Not yet SOC 2 certified — here's what's in place today and what's on the roadmap. RLS on 21+ tables.
  </Card>
</Columns>

## Built by Vaadhlabs

TensorCost is built and operated by **Vaadhlabs**. Reach the team at [support@tensorcost.com](mailto:support@tensorcost.com) — our source repository is private, so email is the right way to reach us for now.