Features
Every feature of a platform installation: what it is, the customer root variable that turns it on, its default, what it needs from you, and which deploy profiles ship its images.
The installation root in infrastructure/environments/customer turns on every feature that works without an input only you can supply. This page lists each feature with its variable, its default, and what it needs. Install in your AWS account lists the variables every installation sets, and Costs lists what each feature costs to run.
On by default
| Feature | Variable | Default | You supply | Notes |
|---|---|---|---|---|
| Identity service | platform_identity_enabled | true | platform_identity_domain_name, platform_identity_email_sender, and its Route 53 zone | The deploy tool always sets it to true. Self-registration stays off; see Users and sign-in. |
| Public API | platform_api_enabled | true | platform_api_domain_name, platform_api_certificate_arn, platform_api_route53_zone_id | Needs the agent runtime, channels, and memory. |
| Management console | console_enabled | true | console_route53_zone_id or a us-east-1 console_certificate_arn | Served on platform_identity_application_origin. |
| Model runtime | always on; bifrost_desired_count | 1 task | Bedrock model access | Hosts model_runtime_hosted_providers, default ["bedrock"]. See Model gateway. |
| Agent runtime and stream gateway | agent_runtime_enabled | true | stream_gateway_domain_name | Creates the run queue, the agent worker, workflows, automations, evaluations, and the model catalog sync. |
| Durable Streams | durable_streams_desired_count | 1 task | none | durable_streams_enabled cannot be turned off. |
| Semantic events | enable_platform_events | true | none | The agent runtime, channels, memory, and event analytics need it. |
| Channels and credential vault | channel_provider_runtime_enabled | true | Provider OAuth apps when you use them | See Channels. Set channel_provider_domain_name to serve webhooks and OAuth callbacks on your own host instead of the execute-api hostname. |
| Memory | memory_runtime_enabled | true | none | Each new organization gets its search index automatically. See Memory and Brain search. |
| Brains | on when the public API and agent runtime are on | on | none | Each organization gets its search index once its memory search is active. Shares the memory embedding model. |
| Agent email | agent_email_enable_runtime, agent_email_activate_receipt_rule_set | true | agent_email_domain, agent_email_route53_zone_id | Activates the account's SES receipt rule set and catch-all receiving, and clears SES account suppression reasons. Needs a dedicated account and an SES receiving region. |
| Lambda MicroVM sandbox | sandbox_lambda_microvm_enabled | true | A Lambda MicroVM region | The deploy tool publishes and verifies the image. Runs code.execute, workspaces, and hosted code. Outside a MicroVM region, set "lambdaMicrovm": false in the deployment target, which turns it off. See Sandbox runtime. |
| Web tools | web_tools_runtime_enabled, agent_runtime_web_tools_enabled | true | none | Runs Firecrawl, its egress proxy, and a Chromium renderer on web-tools.<platform_api_domain_name>, reachable only from the VPC. Switches egress to a NAT instance; see below. |
| Observability history | agent_observability_history_enabled | true | none | Records execution facts in S3 Tables. The hot plane needs it. |
| Observability hot plane | agent_observability_hot_enabled | true | none | Runs ClickHouse on one t4g.medium instance and a 0.25 vCPU task that answers the observability API and console views. The node installs checksum-pinned ClickHouse packages at boot. See Observability hot plane. |
| Event analytics | event_analytics_lifecycle | running | none | Copies every semantic event into an Iceberg table that operators query with Athena through bun run events or an MCP server. Spans all tenants, so it is for operators only. See Event analytics. |
| Data, document extraction, alerts | always on | on | billing_alert_email | DynamoDB, S3, KMS, ops alerts, and billing alerts. |
Web tools and NAT
The in-stack web tools need an egress path that refuses to forward to private and reserved IPv4 ranges. Only the NAT instance mode provides it, so while web_tools_runtime_enabled is true, network_nat_mode defaults to instance: one fck-nat EC2 instance, network_nat_instance_type t4g.small by default, carries all private-subnet egress, including Bedrock and ECR traffic. The root looks up the newest fck-nat AL2023 ARM64 AMI from its publisher at the first apply and keeps it; set network_nat_instance_ami_id to move to another AMI.
A NAT instance costs less than managed NAT gateways but is one instance in one availability zone, so an instance or zone failure stops egress until EC2 recovers it. To keep managed NAT gateways, one per zone, set web_tools_runtime_enabled to false. You can then point agent_runtime_web_tools_enabled at an external Firecrawl with agent_runtime_firecrawl_base_url and agent_runtime_firecrawl_job_ticket_secret_arn, or set it to false too.
Memory and Brain search
Every organization gets working memory search and Brain search with no operator step. Creating an organization emits organization.created, and the memory runtime provisions that organization's index on the installation's OpenSearch domain with memory_runtime_embedding_model at 1,536 dimensions. A provisioning worker creates the index, verifies it with test writes and searches, and activates it. The activation then starts the Brain runtime, which creates the organization's Brain index with the same embedding model, activates Brain search, and queues every Brain document stored before then for indexing. Until the memory index is active, memory search, recall, and writes that update the index fail.
Each step retries on failure. Work that keeps failing lands in the memory-search-placement-dlq, memory-operation-dlq, or brain-search-placement-dlq dead-letter queue, and each raises an ops alert. memory_runtime_reconciliation_scopes is optional: it adds a five-minute drift repair between stored memories and the index for the environments you list, and search works without it.
Every deploy also runs the activateSearchPlacements stage, which activates any organization that still has no index, such as one created before your installation ran this release. Organizations that already have one are left unchanged. bun run memory:search:control and bun run brain:search:control stay available to move an organization's index to another target; functions/memory-runtime/README.md describes them.
Off by default
| Feature | Variable | You supply | Why it is off |
|---|---|---|---|
| Self-service registration | platform_identity_registration_enabled | nothing | Anyone who reaches the identity domain could create an account. |
| Human verification on sign-up | platform_identity_human_verification | A siteverify provider's URLs and site key, and its secret in SSM | It needs an account with a challenge provider. See Users. |
| Browser tools | agent_runtime_browser_allowed_origins | The HTTPS origins agents may open, or set agent_runtime_browser_any_public_origin to true for any public site | With neither, browser tools are not configured and fail closed. |
| Agent-built apps | apps_enabled | apps_domain and apps_route53_zone_id | Apps serve agent-written code on subdomains of a domain you choose. |
| OCR for scanned PDF pages | document_extraction_ocr_enabled | A Region that offers Amazon Textract, and the decisions below | Page images go to Amazon Textract, which bills per page. See OCR for scanned PDF pages. |
OCR for scanned PDF pages
Document extraction reads a PDF's text layer. A scanned page has none, so by default it contributes no text, and a PDF with no readable text at all resolves to no_text. Set document_extraction_ocr_enabled to true to read those pages with the Amazon Textract DetectDocumentText operation.
With OCR on, extraction sends a PDF page to Textract when its text layer holds fewer than 16 visible characters and the page paints an image. Blank pages and pages drawn only with vector shapes are not sent. Extraction sends at most 12 pages per file, the first ones in page order, each at most 10 MB, and at most 4 at a time. Scanned pages past the 12th stay without text, and the file's summary counts them, for example PDF, 40 pages (12 read by OCR, 28 without text). PDFs inside a ZIP file and image files are never sent to Textract. The text of each page that OCR read starts with --- Page N (OCR) ---, so the model and Brain citations show which text came from OCR. Textract reads English, French, German, Italian, Portuguese, and Spanish. See Amazon Textract best practices.
Before you turn it on, check these:
- Region. Textract must run in the installation's Region. The plan reads the Region's Textract endpoint service and fails where there is none. See Amazon Textract endpoints and quotas.
- Quota. Each page is one
DetectDocumentTextcall, and the account's quota is shared by every file read at the same time. The default is 10 transactions per second in us-east-2, 25 in us-east-1 and us-west-2, and 1 in most other Regions. When Textract keeps throttling, extraction reports the file as unavailable rather than storing a result. Completing an upload then answers503 input_file_resolution_unavailable, and Brain indexing retries later. Request an increase before you turn OCR on in a Region with a low default. See Amazon Textract quotas. - Content use. Page images leave the installation for Textract. AWS may use them to improve its AI services unless your AWS Organization applies an AI services opt-out policy that covers Textract. The installation does not set that policy. See AI services opt-out policies.
- Cost. Textract bills each page it reads, and a file that is read again, for example after a retry, is billed again. Inside the VPC, OCR also adds a Textract interface endpoint in each availability zone of the private subnets, billed per hour and per GB. See Amazon Textract pricing and AWS PrivateLink pricing.
Inside the VPC, the extraction function reaches Textract only through that endpoint, and the endpoint's policy admits only the function's role. The endpoint has private DNS on, so inside the VPC the Region's public Textract name resolves to the endpoint. The endpoint admits only the extraction function, so other workloads in that VPC cannot call Textract unless the endpoint's security group and policy admit them. With agent_runtime_enabled set to false, the function runs outside a VPC and calls the Region's public Textract endpoint.
If reading a PDF leaves too little of its time limit for OCR, the file resolves to limits_exceeded. Turning OCR on does not change files the platform has already read. An input file that resolved to no_text stays no_text, and a Brain document version that failed for lack of text stays failed. Upload the file again to read it with OCR.
Service models
The model gateway calls these models outside agent runs. Each must start with a provider prefix in model_runtime_hosted_providers, or the plan fails.
| Variable | Default |
|---|---|
memory_runtime_embedding_model | bedrock/global.cohere.embed-v4:0. Must return 1,536 dimensions. |
memory_runtime_rerank_model | bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0 |
brain_runtime_answer_model | bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0 |
model_gateway_service_fallback_model_ids | ["bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0"] |
web_tools_firecrawl_model_name | bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0 |
web_tools_firecrawl_embedding_model_name | bedrock/global.cohere.embed-v4:0 |
Both default models are also curated, so an environment without a model policy can run them. See Models and model policies.
Profiles and images
| Feature | Image | platform-customer | platform-prod | platform-dev |
|---|---|---|---|---|
| Agent runtime | agent-worker, stream-gateway | yes | yes | yes |
| Model runtime | bifrost, model-gateway | yes | yes | yes |
| Durable Streams | durable-streams | yes | yes | yes |
| Observability hot plane | observability-hot-plane | yes | yes | yes |
| Web tools | firecrawl, web-tools-proxy | yes | yes | yes |
| Agent-built apps | app-gateway | yes | yes | yes |
| Lambda MicroVM sandbox | MicroVM image published by the deploy tool | yes | yes | yes |
platform-prod and platform-dev describe roots kept outside this repository. See Choose a profile.