Operator guide for Valet Dedicated 0.10.1

Limits and known gaps

This page lists what does not work yet, why, and what to do. Each item holds today. The installation is internal and runs in one account.

No public edge

The load balancer is internal, so nothing outside your network reaches the installation. Slack channels and events, inbound webhooks from SaaS products, Composio triggers, and hosted MCP clients do not work.

Use the dashboard, the valet CLI, and MCP clients that run on a machine with network access. See Connect your people.

About 24 orgs

Each org gets its own certificate on the load balancer listener. A listener holds 25 certificates by default, and the apex certificate takes one. An installation serves about 24 orgs.

Ask AWS Support to raise the listener certificate quota before you near it. The request takes days. check warns when 20 certificates are attached. Troubleshoot shows how to count the certificates and what a full listener looks like.

Models

The installation calls Anthropic's API with your key. OpenAI models need your OpenAI key in inference.openai_api_key; without it, the catalog does not offer OpenAI model plans. You can route either provider through your own compatible gateway; see Model gateways. The catalog still follows which key is set, so give a gateway that needs no key a placeholder one. To use Amazon Bedrock instead, see Amazon Bedrock. Otherwise keep the key valid and funded. Change it with settings set inference.anthropic_api_key; see the reference.

Apps keep no state on disk

An app's disk is lost when it restarts. Turn on the app database to give an app PostgreSQL, or keep state in a service you run. Without it, apps have no database resource.

Cold starts

The first request to an idle agent or app waits for a container to start. That takes about a minute. The setting runtime.standby_count keeps warm runtimes ready. Raise it with valet-dedicated settings set.

One availability zone for the database

The database runs in one availability zone, and each platform service runs one task. A zone outage or a task failure interrupts the installation until AWS or the service recovers. Plan maintenance windows around upgrade, which restarts the services it changes.

Teardown leaves resources behind

To remove an installation, set protection = false in the module and apply. Then run terraform destroy. Destroy waits several minutes while AWS releases the load balancer certificate.

Two kinds of resources are not in Terraform's state, so destroy leaves them:

  • Per-org certificates that the platform requested. List them in ACM: certificates for *.<org>.<apex> tagged ManagedBy=valet-api. Delete each one.
  • Task definitions that upgrade registered. List them in ECS: the families named <name>-*, where <name> is the module's name input. Deregister every revision, then delete them.

Remove the leftovers after destroy finishes. A certificate still attached to the listener blocks deletion until destroy detaches it.

Access rules

An operator can forbid public or password access across the installation. Org Admins can add stricter rules in Team settings. A new rule does not change existing services. The dashboard and valet-dedicated access-rules get list services whose current mode conflicts with the rule. Change those services to private access when you are ready. See the command reference.

Custom connectors

An org Admin adds custom connectors by default. An operator can instead allow Developers or restrict definitions to approved hosts and npm packages. Existing connectors and catalog installations remain available. See the custom connector reference.