Secrets and model access keys

Overview

Two small primitives keep credentials out of places they do not belong:

Secrets are workspace-scoped named values (OPENAI_API_KEY, GITHUB_TOKEN, …) stored encrypted. The API never returns a value once stored. Agent environments reference secrets by name.
Model access keys are API keys bound to one deployment. They authenticate only that deployment’s inference routes, so you can give one to an agent sandbox, a coding assistant or another hosting provider without exposing your workspace.

Secrets

Names follow environment-variable syntax ([A-Za-z_][A-Za-z0-9_]{0,63}), because that is how agent sessions receive them.

# Prefer --from-env or --stdin so the value never lands in shell history.
export OPENAI_API_KEY=sk-...
veri secrets set OPENAI_API_KEY --from-env OPENAI_API_KEY --scope-host api.openai.com

veri secrets list
veri secrets delete OPENAI_API_KEY

from veri_sdk import Client

client = Client()
client.secrets.set("GITHUB_TOKEN", token)                       # injected as an env var
client.secrets.set("OPENAI_API_KEY", key, scope_host="api.openai.com")
client.secrets.list()   # names and metadata only

curl -X PUT https://api.veri.studio/v1/secrets/OPENAI_API_KEY \
  -H "Authorization: Bearer $VERI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"value": "sk-...", "scope_host": "api.openai.com"}'

Field	Meaning
`value`	The secret. At most 64 KiB. Encrypted at rest, never returned.
`scope_host`	Optional bare hostname (`api.openai.com`). When set, agent sessions attach the value only to HTTPS requests for that host, and the value itself never enters the sandbox. When absent, the value is injected as an environment variable named after the secret.

PUT on an existing name replaces the value in place. Secrets are visible only inside the workspace that created them.

Model access keys

A normal API key can do everything in your workspace. A model access key can do exactly four things, all on one deployment:

Method	Route
`GET`	`/v1/deployments/{id}`
`POST`	`/v1/deployments/{id}/chat/completions`
`POST`	`/v1/deployments/{id}/messages`
`POST`	`/v1/deployments/{id}/messages/count_tokens`

Everything else, including other deployments, returns 401.

veri api-keys create agent-prod --deployment dep1a2b3c4d5e --expires-in 86400
# prints the key once

veri api-keys list
veri api-keys revoke <key_id>

key = client.api_keys.create("agent-prod", deployment_id="dep1a2b3c4d5e", expires_in_seconds=86400)
key["key"]      # plaintext, returned once
key["kind"]     # "model_access"

Point any OpenAI-compatible client at the deployment with this key:

OPENAI_BASE_URL=https://api.veri.studio/v1/deployments/dep1a2b3c4d5e \
OPENAI_API_KEY=vk_... \
your-agent

Field	Meaning
`deployment_id`	Binds the key to one deployment. The deployment must belong to the workspace minting the key.
`expires_in_seconds`	Optional lifetime, 60 seconds to one year. An expired key is rejected exactly like a revoked one.
`kind`	`user` for a full key, `model_access` for a deployment-bound key.

Managed agent sessions mint a short-lived model access key per session automatically when an environment’s model is a Veri deployment.

Get started

Training

Deployments

Evaluations

CLI

Volumes

Webhooks

Secrets

Billing

Overview

Secrets

Model access keys

​Overview

​Secrets

​Model access keys

Overview

Secrets

Model access keys