Admin Settings

A few settings belong to the installation rather than to an integration or a person: how it sends email, which model its agent reasons with. They live in the admin section, reached from Admin in the account menu (the avatar, top right) at /platform/admin.

Prerequisite: an encryption key

Provider API keys are encrypted at rest with the same AES-GCM cipher the KV secret namespaces use, keyed by KV_ENCRYPTION_KEY. The chart leaves it unset by default, so on most installs it has to be turned on:

kv:
  encryptionKey: "<base64 of 32 random bytes>"

Until it is set, both admin pages show an amber banner, the API-key field is disabled, and an attempt to store a key is refused with a 503 naming the variable. A key is never stored in the clear. Everything else on those pages still saves without it, because carrying an already-stored key forward copies ciphertext and decrypts nothing.

Email

/platform/admin/email configures outbound email, sent through Resend.

The from address is rejected if it carries a display name (Acme <a@b.co>); that belongs in the From name field.

Send a test

The Send a test panel uses whatever is in the form right now, including a key you have not saved yet, so you can paste a key, send yourself a test, then save. With no key typed it falls back to the stored one, so the button still works after a reload. Failures say what to do: an invalid key says so, and an unverified domain comes back in Resend's own words, naming the domain.

Sending from your own code

Once configured, anything running inside the cluster can send through the site's identity:

curl -sS -X POST "$ORCHESTRATOR_URL/email/send" \
  -H "Authorization: Bearer $TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"to":["you@example.com"],"subject":"hello","text":"body"}'

Sending needs a token holding operator or admin. A deployed integration presents its own — env.PLATFORM_TOKEN, see platform access — and the operates deployments grant is what carries it, so a flow that has just repaired something can report that it did. Reading or changing the settings above stays with administrators: those hold the API key, and sending through a server somebody else configured does not.

The endpoint takes no API key and no from-address: both come from the stored settings, and supplying either is rejected as an unknown field. Either text or html (or both) is required, and a request may carry at most 50 recipients; this is a transactional sender for notifications, not a bulk mailer. It returns {"messageId": "..."} on success, and 409 until email is configured.

Platform agent

/platform/admin/agent is one page for one task: the model the platform's own agent reasons with, and his deployment. He will not install without an LLM provider, so the page that refuses is the page holding the field.

Three sections, in the order they are needed: the LLM provider first, Web search second, the Deployment third. Embeddings are not here: they configure how agent memory is searched rather than how the agent reasons, and what is left to say about them is reported where the searching happens.

LLM provider

Model is free text because model names turn over faster than releases of this app.

OpenRouter is the one to pick if you want to try several models. It fronts hundreds of them behind a single key, so changing which model Dr. Octo reasons with is a change to the model field and nothing else. Its ids carry the vendor as a prefix (anthropic/claude-sonnet-4.5, openai/gpt-5.4, google/gemini-3.5-flash), and there is no bare model name. See the connector reference.

Web search

One field: a Parallel API key. With one, Dr. Octo holds a web_search tool that searches the open web and hands back ranked pages with the excerpts that bear on what he asked.

It is optional. Without it the tool answers that web search is not configured, tells him not to call it again, and he says he could not check the web. The key reaches his pod as an environment variable, and when there is none the value is a sentinel his tool compares against before calling anything.

The key is write-only, encrypted at rest, and reported as its last four characters. It is bound onto his deployment as a cluster secret, and bindings are written on install and roll-out, so a key saved while he is running is picked up on his next Roll out.

Deployment

The third section installs, updates, traces or removes him, sets his turn limit (how many tool-calling turns one answer may take) and holds the checkbox that lets his troubleshooter act on alerts. The page has one Save at its foot: it writes the provider first, then the search key, then the two pod-level settings in a single call so his pods are replaced once. The button says so when a pod-level field is dirty. See The Platform Agent for what he does and how to change him.

Embeddings (optional)

Whether searching agent memory ranks by meaning or by words. Leave it out and search matches text, using Postgres full-text. Deploy the embedding server and the same search ranks by embedding similarity instead, with the same query and the same results shape.

The embedding server

Embeddings are produced by a small octo app the chart deploys: embeddings/config.yaml, an ai-embed block behind an HTTP route. It is a service because it holds the provider API key, so that no integration pod has to. Every deployment pod is given its address as EMBEDDINGS_URL, so any integration can embed text; an embedding reads nothing, writes nothing and costs a fraction of a cent. (Contrast OBSERVABILITY_URL, which is a grant: stored telemetry is other integrations' data.)

Turn it on in the Helm values:

embeddings:
  enabled: true
  connectorType: llm-openai   # or llm-gemini, llm-openrouter
  model: text-embedding-3-small
  dimensions: 1536
  apiKey: "sk-..."            # or existingSecret

connectorType has no Anthropic option: Anthropic has no embeddings API, so the connector does not implement the capability and the flow is refused when it is built, as a startup failure naming the problem.

dimensions must match the vector(N) columns in sql/schema.sql, which are indexed and therefore fixed at schema time. It is requested of the provider, and a model that cannot produce that width fails the call and says so.

On a Terraform deploy the same settings are embeddings_* variables on the release root. The key is persisted to the state bucket as release/embeddings.json, the same mechanism the OIDC credentials use, so a Cloud Build deploy (which has no terraform.tfvars) reads it back instead of tearing the server down.

The embedding server appears on Platform services like any other dependency. "Not configured" there is the ordinary answer.

The backfill

Everything written before the server existed has no vector, so a background sweep in the orchestrator works through them, newest first. The agent memory page shows how much is left. While the sweep runs, a search may be answered by either index: the vector one when it has something to match, the text one when it does not. The sweep is paced at 64 rows every 5 seconds, so a history of hundreds of thousands of items is not sent to a paid API in one burst.

How a save handles the API key

Leaving the key field blank does not clear the stored key. The three cases:

Both forms submit every field on every save, so any other design would mean editing a from-address destroyed the key beside it. Saves take a row lock, so two people saving at once cannot silently discard each other's key rotation.

Storage

The email, LLM and web search settings are rows in the site_settings table (key = 'email', 'llm' and 'websearch'), each a jsonb document. Embeddings are not among them, because nothing about them is a setting. Adding a setting costs a key, not a migration, which is how data retention below was added without one.

What agents remember is not configuration and is not here: it has its own tables and its own section of the platform. See Agent Memory.

Data retention

Set how long this installation keeps stored logs, traces and alert history at /platform/admin/retention. A ten-block flow emits a couple of dozen trace records, each able to carry a captured request body, where the same request produces a log line or two, so give traces the shorter window.

Logs and traces default to 0, so an installation that never opens this page deletes nothing until you say so. Each field says beneath it what its value means, so a bare 0 reads as a decision rather than a value nobody set.

Saving deletes nothing. By default a nightly job applies the policy at 03:00; set retention.schedule in the Helm values to change when it runs, or retention.enabled=false to remove it.

Run now

Press Delete now to apply the saved policy immediately instead of waiting for tonight. It asks first, naming the windows it is about to enforce, then reports what went: log events, trace records, traces, and alert evaluations.

Save before you run it; the button stays disabled while the form has unsaved edits, because a sweep enforces what is stored.

Two behaviours are explained in Traces: a trace is deleted whole rather than by record, and stored model prices are never deleted, because a traced call's cost points into that history.

Platform services

/platform/admin/health reports rather than configures. It asks the orchestrator whether it can currently reach the five things this installation may run on:

Each check is one round trip: it proves a dependency answered and no more. Postgres answering says nothing about replication lag, and Redis answering says nothing about how full it is.

There are three results. Not configured is neutral rather than a warning, because running an orchestrator without cluster access, or without embeddings, is a supported way to run. Unreachable shows the transport error verbatim, the string you would otherwise go to the pod logs for. It checks on load and on Check again, and does not poll.