Skip to content

Web console overview

The web console is part of Prism: prism serve starts it on its own listener, 127.0.0.1:18081 by default, and nothing else has to be installed. The page is http://127.0.0.1:18081/prism/ui/. The console listener serves only the console and its API; model requests go to the main address (127.0.0.1:18000).

The console is on when [personal] enabled is true, which is the default. Set [personal] enabled = false for a Prism without a console and without user accounts.

There are three ways in:

Way Who Needs
Sign in with owner secret the owner the console token or PRISM_ADMIN_TOKEN
Sign in with a passkey anyone with an account a passkey added earlier; HTTPS, or the console opened as http://localhost:18081
Email link anyone with an account mail configured in Configuration → Email

Owner secret. On first start, if PRISM_ADMIN_TOKEN is not set, Prism creates a token in <storage.dir>/console-token (mode 0600) and uses it on every restart. PRISM_ADMIN_TOKEN in Prism’s environment replaces the file. prism console prints the URL and the token:

18081/prism/ui/
cd ~/.prism && prism console
# token: 6f1c…

The token stays in that browser tab until you sign out.

Passkey. After signing in, add a passkey from the profile menu. Passkeys are bound to the console’s host name: open a loopback console as http://localhost:18081/prism/ui/, not 127.0.0.1 (an IP address cannot hold passkeys). For a console on another address, see Open the console from another machine.

Email link. With mail set up, the sign-in page sends a one-time link to the account’s address. The link works once and for 15 minutes.

A session ends after 12 hours without use or 30 days after sign-in, whichever comes first ([personal] session_idle_timeout, session_max_lifetime).

The owner invites people under Users and budgets → Members. An invitation link works once and for 72 hours. Prism Free allows three people, the owner included; Pro allows up to 5, Teams up to 10 and Enterprise its paid seats. A pending invitation, agents and API keys take no seat. A signed-in user who is not the owner sees only their own connections, keys, requests and profile.

Menu item What it is for
Overview spend for the period, quotas of subscription accounts, subscription prices, client addresses
AI providers connect API keys, OAuth subscriptions and your own endpoints; test a connection; models of each connection, with a search; who may use it; keys from 1Password
Compute providers model servers and machines Prism starts and stops (Pro and above); see Managed inference; local model servers (the former [[local]] entries), while there are any
OAuth providers sign in to subscription accounts, their quotas; the Auth files tab: sign-in files of subscriptions, refresh, rename, delete, import a file from Codex CLI, Claude Code or Gemini CLI
Routing default strategy and model, seller chain, account selection, routing evaluation, tool-output compression
Agents traffic, cost and models per client application; sessions
Users and budgets members and invitations, roles, spending limits, provider quotas
Logs find a trace by ID
Usage history spend over time, the cost of your own model machines and the machine budget (machine cost), analysis by provider, model and application, the request journal (Request details), prices
Configuration settings saved to the console overlay, API keys for clients, learned rules, operator tools, restart and stop
Diagnostics the status snapshot, the capability manifest, a routing preview for a request
Plugins install and manage extension packages, their settings and panels (owner only); see Installed packages
Versions installed and latest versions, updates, release components
Quick start client addresses and ready-made client settings, agent skills, a media workspace

The search field (⌘K or Ctrl+K) finds sections, settings and actions by name.

OpenAI’s Usage Policies (from 2026-11-20) ask a service that gives people its models to name the person behind each request and to act on that one person when OpenAI reports a violation. In Teams and Enterprise Prism does this:

  • SafetyID. Every request that reaches an OpenAI model carries a safety_identifier for its person or service account: prism-sid- and a hash that carries no name, address or ID. A client’s own value is replaced.
  • Refusals are not passed on. When an OpenAI model refuses a request, Prism does not send it to another OpenAI route (another account, a reseller); other companies’ models may still answer it.
  • Finding the person. Organization → Activity log has a lookup for the owner and administrators: paste the SafetyID from OpenAI’s message to see the person (address, team, OpenAI refusals over 30 days) or the service account. Each lookup goes into the activity log; the person found does not see it.
  • Refusals per person. The members table has an OpenAI refusals column for the period shown, and the person card shows the last 30 days.
  • Suspension. On the person card the owner or an administrator can suspend a person: Prism then sends none of their requests to OpenAI models, by any of their keys, while other companies’ models keep serving them. A request only OpenAI models could serve is refused with 403 openai_access_suspended. Keys, role, team and history stay, the person can still sign in, and the suspension can be lifted at any time. Both are recorded in the activity log.

Users and budgets → Budgets edits the spending limits (configuration reference). Each limit has a period (hour, day, week, month), an amount in US dollars, an optional app (only requests with that X-Prism-App count) and the point where the period starts, in UTC. Its thresholds say what Prism does at a share of the amount: a warning header, cheaper models, local models only, or refusing requests with 429. The list is saved whole and applies after a restart; while the saved limits are the running ones, each shows what has been spent in the current period.

Above the models of a connection on AI providers is a search field. Type what you need, a name or a description such as “cheap, for code, reads images”. Models whose name, description or tier contain every word appear at once, with the matches in the list. Prism then adds models close in meaning, marked by meaning: a local embeddings model ranks them by what their row says and by their price among the connection’s models. The query and the model descriptions do not leave your machines; Prism uses the embeddings model that semantic memory uses (Experiments → Local embeddings) and never a cloud one. The first search reads every model of the connection once, which on a CPU can take a minute or two for several hundred models; the line under the field shows how far it got, and later searches are immediate. Without a local embeddings model the search works by words, and the line links to the card that installs one. While a query is typed the list is sorted by match; it can also be sorted by name, price or speed.

Settings are changed in the console. On its first start after the update, Prism moves the settings from config.toml into <storage.dir>/settings.toml once and keeps a copy of the old file; keys left in config.toml then take no effect, except the few Prism needs to start. Configuration, Routing, AI providers, Compute providers and Users and budgets → Budgets write to that store. Webhook, fleet and Laya keys are stored write-only in a sealed vault and are never shown again; without a stored value, the environment variable the configuration names still works. Prism reads its configuration once: a saved value takes effect at the next restart. Restart Prism is available when Allow restart and stop from the console is on (Configuration → Maintenance); see Update Prism.

Configuration → Experiments turns on features that run beside routing and change no answer: the model choice check, background answer review, secret detection, request skill detection and user reaction scoring with Laya, semantic route memory and local embeddings. Each has a switch and the values it needs.

Each experiment that collects something has a goal on its card on the Experiments overview: how much it has collected of the goal, and when it will get there at the last week’s pace. Change the goal on the card. When a 50/50 test reaches its goal it stops, new conversations go back to the ordinary choice, and its result stays on the card; raise the goal to continue it. A background experiment that reaches its goal keeps recording until you switch it off. Prism tells you in the notification bell (category Experiments, also by email if you choose it) when an experiment reaches its goal or a safety rule stops it, so you do not have to check. Local embeddings collect nothing and have no goal.

Each card also says which decision the experiment serves and what to do once it reaches its goal, and holds a Task field for your own link or tracker number; the task is shown on the card and in the experiment’s notifications.

The console is available in English, Russian, Chinese (Simplified), Spanish, German, French, Ukrainian, Hebrew and Italian. Each person chooses the language in the profile menu; the choice is saved with the account.