Costs & AWS Resources

What the CDK stack creates in your account and what running the platform infrastructure costs.

← Back to README

AWS Resources Created

The CDK stack (infra/stacks/platform_stack.py) creates:

• API Gateway HTTP API -- Routes /api/workflows/* to Workflow Lambda, /api/deploy, /api/test-runtime, /api/runtime/*, /api/generate-tool to Deployment Lambda. HTTPS by default, CORS configured.
• Workflow Lambda -- FastAPI app wrapped with Mangum. Handles workflow CRUD, validation, import/export.
• Deployment Lambda -- Handles deploy initiation, status polling, runtime testing, runtime deletion (full cleanup), AI tool generation via Claude Sonnet on Bedrock, and CloudFormation template generation/export. 120s timeout for LLM calls.
• Step Functions State Machine -- Orchestrates multi-step deployments: validate -> [mcp_server?] -> [knowledge_base?] -> [gateway?] -> [memory?] -> [policy?] -> codegen -> IAM -> runtime configure -> runtime launch -> [evaluation?] -> [auth?] -> status update. 3 retries with exponential backoff per step.
• Step Lambdas -- Individual Lambda functions for each deployment step (validate, codegen, IAM, gateway, knowledge_base, mcp_server, memory, policy, evaluation, runtime_configure, runtime_launch, auth, status_update).
• DynamoDB Tables -- Workflows + Deployments (TTL + GSI on workflow_id/runtime_id), plus the enterprise-feature stores: agent-versions, runtime-slots, agent-registry, prompt-library, hitl-requests (24h TTL), triggers, usage-events (90d TTL), budget (persistent definitions, 400d monthly counters, 90d settled reservations), audit (90d TTL in development; no TTL in production), permission-requests, the shared org-config table (tag policies + VPC profiles + approval policies), and flows. Each user-data table carries an owner_sub GSI for owner-scoped list queries.
• Cognito User Pool Groups -- registry-admin / registry-developer for the registry two-persona approval model, plus the RBAC groups t-admin/t-user (tenant role) and g-admins-*/g-users-* (resource scopes). See Registry & RBAC.
• EventBridge Rules -- policy-sweep (rate 5 min — converges pending Cedar ENFORCE permits to ACTIVE) and cost-reconcile (rate 24 h — commits unknown stale invocation reservations conservatively, then evaluates month-to-date estimated-cost thresholds). Both invoke the Deployment Lambda with a sentinel payload.
• S3 Bucket -- Frontend static assets (CloudFront OAI access only) + AgentCore dependency bundles + deployment code artifacts.
• CloudFront Distribution -- HTTPS, SPA routing (404/403 -> index.html), API Gateway as additional origin for /api/*.
• SSM Parameters -- CORS origins, AWS region, DynamoDB table name under /agentcore-workflow/{env}/.
• IAM Roles -- Least-privilege per function: Workflow Lambda gets DynamoDB workflows + SSM read; Deployment and long-running stream Lambdas get budget definition reads, conditional counter/reservation writes, and dynamodb:TransactWriteItems only on the budget table; Deployment Lambda also gets deployments + bedrock-agentcore + bedrock:ListFoundationModels/ListInferenceProfiles (live model catalog) + cloudwatch:PutMetricData (budget breach) + JIT iam:PutRolePolicy scoped to AgentCore* roles + cleanup permissions (Cognito, Lambda, STS); Step Lambdas get deployment permissions required by their exact phase.
• CloudWatch Log Groups -- Lambda and Step Functions execution logs.

All resources are tagged with environment and project for cost tracking.

Infrastructure Pricing

This section covers the cost of running the platform infrastructure itself (API Gateway, Lambda, Step Functions, DynamoDB, S3, CloudFront, etc.). It does not include AgentCore deployment costs (Runtime compute, Gateway invocations, Memory storage, Bedrock model inference) which are billed separately by AWS Bedrock AgentCore.

All prices are for us-east-1 (N. Virginia) as of 2025. See the linked AWS pricing pages for the latest rates.

Per-Service Breakdown

Monthly Cost Estimates

Estimates assume all agent deployment activity flows through Step Functions (the primary path from the UI). Each deployment triggers ~16 state transitions and ~16 Lambda invocations.

With the AWS Free Tier (first 12 months or always-free tiers), this platform costs effectively $0.02/month at low usage and under $0.40/month at moderate usage. Even without any free tier, moderate usage stays under $2/month. The serverless, pay-per-request architecture means you pay nothing when the platform is idle.

What's NOT Included

The table above covers platform infrastructure only. The following are billed separately by AWS and depend on your agent usage:

• Amazon Bedrock model inference — per-token pricing varies by model (e.g., Claude Sonnet, Nova)
• Bedrock Guardrails — production tool calls add one ApplyGuardrail input screen before execution and one output screen before model consumption; policy-unit charges vary by configured filters. A policy node with a Gateway guardrails profile adds one InvokeGuardrailChecks evaluation per Gateway tool call (prompt-attack under standard; prompt-attack plus content filters under strict), billed to the Gateway role's account
• AgentCore Runtime compute — billed per-second while the runtime is active
• AgentCore Gateway invocations — per-request to the MCP gateway
• AgentCore Memory storage — per-GB for conversation memory
• Bedrock Knowledge Base — per-query for RetrieveAndGenerate, vector store costs (OpenSearch Serverless / RDS Aurora), S3 storage for source documents
• AgentCore Code Interpreter / Browser — per-session usage
• Cognito — first 50,000 MAUs are free; OAuth2 machine-to-machine auth adds $0.0065/token request beyond 250/mo
• Custom tool Lambda functions — same Lambda pricing as above, billed per invocation

Refer to the Amazon Bedrock AgentCore pricing page (opens in new tab) for current rates.

Invocation authorization versus actual spend

Production invocation routes require an applicable owner/agent/tag limit. For one top-level operation (including a bounded MCP handshake plus tool call), one DynamoDB transaction reserves the configured budget_reservation_usd unit across every matching limit before the first Runtime dispatch; a concurrent request cannot push an aggregate counter above the current definition. Exhaustion returns HTTP 429 with Retry-After: 3600; storage uncertainty returns 503 before dispatch. Any request that reached the Runtime network boundary commits its unit, even if Runtime returned an error. Calls proven not dispatched release it, while unknown stale reservations commit after one hour.

These USD-denominated units are admission capacity, not metered charges and not a bill guarantee. The actual-cost panel separately prices observed Bedrock token telemetry, and AWS billing can also include Runtime duration, Gateway, Memory, Browser, Code Interpreter, Knowledge Bases, custom tools, and third-party vendor charges. Calls made with separately granted AWS or SaaS credentials are outside the ledger. Production A2A is therefore refused until inbound and nested peer calls can propagate reservations; OBO Runtime remains supported because only the platform's SigV4 role can invoke it after reservation. Use AWS Budgets/Cost Anomaly Detection, service quotas, least-privilege IAM, and vendor-side limits for account-level spend control.