---
title: "Prism FAQ: Self-Hosted AI Gateway"
description: "Answers about Prism: setup, providers and local models, subscriptions and failover, cost, privacy, and the Free, Pro, Teams and Enterprise editions."
url: https://prism.tm/faq/
---

FAQ

# Questions about Prism.

Everything you need to know about setup, model routing, subscription accounts, local runtimes, cost tracking, and privacy in Prism.

**Prism is not yet publicly available.** [Request access](https://prism.tm/request-access/)  

## About Prism

### What is Prism?

Prism is a self-hosted adaptive AI gateway. Your agents and apps call one API, and on every request Prism chooses the model, provider and account to use: a paid API, one of your subscription accounts, or a model on your own hardware. Cost, access and diagnostics are managed in one console.

### Who is Prism for?

Developers and teams who run several AI agents, coding tools and scripts and do not want to configure every provider in each of them. Prism Free costs nothing for any use, by you, your team or your company, with up to three people in the console. Pro adds managed inference and Smart routing for up to five people, Teams adds an organization with roles for up to ten people, and Enterprise adds several teams.

### Can I try Prism today?

Not yet. Prism is not yet publicly available, and access is on request. [Request access](https://prism.tm/request-access/)

### Is Prism open source?

No. Prism is proprietary software, but Prism Free costs nothing and needs no licence.

## Setup

### What do I need to run it?

A Linux or macOS machine, x86-64 or Arm — that is it. Prism is a single Go binary with the web console and storage built in: no Docker, no Python, no database server, and no root. A one-command installer sets it up as a systemd service on Linux or a launchd service on macOS, and it listens on localhost by default.

### How do I connect my agents and apps?

Point them at Prism instead of the provider. Prism serves an OpenAI-compatible API — chat completions, Responses, embeddings, rerank and models — plus native Anthropic Messages and Gemini endpoints, so most tools need only a new base URL and a Prism key.

Apps that speak only Ollama, or offer an Ollama provider option — Open WebUI, Zed, Continue and Home Assistant — can point at Prism's main address and reach every model Prism routes; the console's quick start shows the address. If an app has no field for a key, put a person's Prism key in the address: `http://host:18000/k/<key>`. Prism Teams requires a Prism key on every model request.

Each person and each agent can have a separate Prism key, which you can issue, rotate and revoke. A person gets their key in one click under My Prism key in the console, and requests made with it are recorded under that person's name.

### Does Prism work with Codex and Claude Code?

Yes, with your API keys. Point Codex at Prism and it can use Claude, Gemini and other models, not only those served through the Responses API. Codex's reasoning effort and fast mode settings also work with a Prism key. Claude Code connects through Prism's Anthropic Messages endpoint, with its web search working. Claude Code can also use models other than Claude. Both are named for compatibility only — Prism is not affiliated with OpenAI or Anthropic.

### Which AI providers does Prism support?

Most connect with just an API key, and Prism discovers their models automatically: OpenAI, Anthropic, Google Gemini, Meta Model API, OpenRouter, Groq, DeepSeek, Together AI, Fireworks AI, DeepInfra and more, plus any OpenAI-compatible endpoint. The console also lists the providers of the public models.dev catalog to pick from.

Amazon Bedrock (Claude included), Google Vertex AI (Gemini and Claude models) and Azure OpenAI / Foundry connect through plugin packages that you install from the console's Plugins page. Cloudflare Workers AI, Z.ai, MiniMax, Novita AI, Parasail, SambaNova Cloud and Vercel AI Gateway are available the same way. The Meta Muse Code package adds Meta's Muse Spark and Muse Wonder models; you connect it by signing in with your Meta account.

### Can I use local models?

Yes. Add the URL of an Ollama, llama.cpp, vLLM or LM Studio server and Prism finds its models. Local models route under the same policy as paid APIs, and requests you mark private never leave your local runtimes.

### Can I connect a model server I already run?

Yes. Paste its address, such as `http://host:port/v1`, and Prism asks the server which API it speaks, names the connection after it, and connects without a key when the server checks none — asking for a key only when the server wants one. This is part of Prism Free.

### Can Prism start and stop my model servers?

Yes, with Prism Pro or a higher edition. On the Compute providers page, the Add machine wizard guides you step by step. For the machine Prism runs on, no command is needed when its model servers run under Prism's own account; for another Linux machine, you run one command as the user that owns its model servers. A new cloud instance can also connect itself: paste the text the wizard shows into the instance's user data, and no terminal is needed. In each case a small service agent is installed, and Prism does nothing on the machine until you confirm the machine in the console. The agent finds llama-server, vLLM, Ollama, SGLang and any other server that answers `/v1/models`, including servers that run as system services under another account (those need a one-time grant on the machine), and you tick the ones Prism manages.

When a request names a stopped model, it waits while Prism starts the server and checks it is ready, and that model answers it; when Prism chooses the model itself, the request goes to another model straight away. A stop lets running requests finish first, and a server can stop on its own after a quiet period. Connecting to model servers that are already running is included in Prism Free.

### Can Prism use GPU machines in the cloud?

Yes, existing machines in your cloud account (Pro and higher editions). Support for Nebius is built in; Amazon EC2, Google Compute Engine, Microsoft Azure, Oracle Cloud, Alibaba Cloud ECS, Scaleway, Crusoe, Hyperstack, Paperspace and Oblivus are added with plugin packages. Add the cloud account's credentials and pick a machine, and Prism can power it on and off. A request that names a model on that machine powers it on, waits for the agent, starts the model server and is answered by that model; requests that arrive meanwhile share the same power-on.

Prism powers the machine off after the idle time you set, but not while other processes are using the GPU, and the console shows what is running. Prism only starts and stops machines; it never creates, resizes or deletes them.

### Can Prism power on a GPU machine on my own network?

Yes, with Prism Pro or a higher edition, through Wake-on-LAN. When a request needs the machine's model, another machine with Prism's agent in the same subnet sends the wake-up packet. Prism powers the machine off through its own agent after stopping its model servers, either when you press Power off or after the idle time you set. Wake-on-LAN must be turned on in the machine's BIOS or UEFI settings, the machine needs a wired connection, and power-off needs a one-time permission on the machine.

### Do I have to edit config files or restart Prism to add a machine?

No. Machines, model servers and cloud accounts are added on the Compute providers page and take effect without a restart. Agent tokens and cloud keys sit in Prism's sealed vault, and Prism refuses changes to a server that is busy with requests.

### Can my agents manage model servers and machines?

Yes. Create an API key with Prism access for an agent and choose what it may do: add, check and remove its model servers and see machines and model services, and with Prism Pro or a higher edition also start, stop and restart services, power linked cloud machines on and off, and ask a machine to find its servers again. The agent works through Prism's API instead of SSH, and every action is recorded with the key's name. Such a key cannot call models and opens no keys, users or licence.

### How do updates work?

From the console, or with `prism update install`. A release updates as a set — Prism and the companions installed beside it — and if the new version fails to start or pass its health check, everything rolls back by itself. Prism checks each update's signature and can also update offline. Prism snapshots its state before each update, and a built-in backup command copies the rest.

### Can I add providers and integrations without upgrading Prism?

Usually, yes. On the console's Plugins page, install a plugin package from the catalog at [prism.tm/plugins](https://prism.tm/plugins/); a package that needs a newer Prism version says which one. Packages add AI providers, cloud accounts for managed inference (Pro and higher editions), OpenTelemetry export and spend alerts. Prism checks each package's signature before it enables it. On a network without internet access, download the package file from prism.tm/plugins and choose Install file on the Plugins page.

## Subscriptions and routing

### Can Prism use the subscriptions I already pay for?

Yes. Prism signs in to many kinds of supported subscription account — chat and coding plans alike — several accounts of each, and routes them alongside your API keys and local models; when an account hits an error, a rate limit or its quota, Prism moves to the next eligible account and reports why.

Sign-in to Claude, Gemini, Cursor and Grok CLI subscriptions is closed by default because these providers' terms forbid third-party programs, and opens only if the Prism team allows it for your installation. Sign-in to Codex, GitHub Copilot, Kimi Code, Kilo Code, Cline, iFlow, xAI, Kimchi and CodeBuddy is open by default.

Two caveats: operation on live accounts is not yet confirmed (it is tested against test servers), and using a subscription through a third-party gateway may breach the provider's terms. Before every sign-in Prism shows a risk notice that says how Prism identifies itself to the provider.

### How does Prism choose where a request goes?

On every request, Prism picks the model, provider and account from your policy and live signals: capability, health, quota, remaining funds and cost. With `auto`, each request gets a quality floor — higher for agent steps that use tools — and the strongest model at or above it serves, with price deciding only between equals and response headers saying why. Pin a model as `provider::model` to keep the request on that provider unless you allow cross-provider failover, or use a selector: `auto`, `fast`, `smart`, `cheap` or `local`.

When Prism chooses the model and that model is available through several routes, it tries them in the order you choose.

### What happens when a provider fails?

Prism sends the request to the next eligible provider or account. This happens on errors, rate limits, timeouts, outages, exhausted funds and content-filter refusals that come before any output; when Prism chose the model itself, a model's refusal (HTTP 403, other than an authentication or billing error) also moves the request on. When a provider refuses a request because the model lacks something the request needs, such as image input, the request moves on. This also happens when the request named the model; it then goes to another provider or account that serves that model. An answer that ends abnormally before anything has reached your app, such as a Gemini answer stopped by a malformed tool call or a Responses answer with the status failed, also moves the request on. Response headers name the provider that served the request and list each attempt with its reason. Providers that run out of funds are set aside and checked again automatically.

### Can I keep some requests strictly local?

Yes. Requests marked private go only to your local runtimes and never fall back to a cloud provider — if no local model can serve them, they fail instead.

## Cost

### What does Prism cost?

Prism Free costs nothing, for up to 3 people, with no limit on your agents, keys or provider accounts. The paid editions are priced per user per month, before tax: Pro €4, Teams €6, Enterprise from €8 with at least 5 seats. A year costs ten months: €40, €60 and from €80, before tax. Tax is added on top. Every person in the installation is paid for, the first three included; agents and keys never take a seat. The paid editions are not on sale yet. Your provider usage is billed by your providers, as before.

### Does Prism take a cut of my provider spend?

No. Prism adds nothing to provider prices and takes no share of your usage.

### How do I see what I spend?

Prism records the cost and tokens of each request. When the provider reports what it charged, Prism records that amount; otherwise it calculates the cost from the price list, with cached-token and long-context rates and, for a request served in the priority or flex service tier (such as Codex fast mode), that tier's rates. The console totals cost by app, client program, agent, person and model, and shows metered API spend apart from what your subscription traffic would have cost at API prices, so you can see what your subscriptions save.

### Can I set budgets?

Yes. Budgets are opt-in and set in the console. A budget covers an hour, day, week or month and applies to everyone, one person, one app, or one person in one app. At percentages of the budget that you choose (for example 80% and 100%), Prism can warn, switch routing to cheaper models, use only local models, or reject requests.

Separate spending limits can cap one API-key provider account, or one model across every provider that offers it. An account that reaches its limit takes no requests until its period resets. When a model reaches its limit, `auto` uses another model, and a request that names that model is refused.

A plugin package for spend alerts can also send a Telegram or Slack message when the day's spend reaches the percentages of a daily budget that you choose in the package, and a summary of the previous day.

If Prism stops unexpectedly, requests that were still running may be recorded as incomplete or not recorded at all. Their cost is not counted, so actual spending can go over a budget by that amount.

## Privacy and security

### Do my prompts go through a Prism cloud?

No. Prism runs on your machine or server and sends requests straight to your providers, not through any Prism server. There is no telemetry unless you turn it on. By default Prism downloads a few public files, and these downloads send nothing about your installation: the model price list (LiteLLM's public file on GitHub) every 24 hours, the model catalog and subscription plan prices from releases.prism.tm once a day, and the plugin catalog from prism.tm every hour. On a closed network you can switch these downloads off.

### Where are my keys stored?

With you: in Prism's encrypted store, in environment variables, or read from 1Password. When you add a provider in the console, Prism sends its key only to that provider's own address. Subscription tokens are stored encrypted and refreshed automatically.

### What does Prism log?

A request journal without prompt bodies, with CSV export, that shows for each request the requested model, the model that served it, the reason for the choice, the number of provider attempts and the trace ID. Also usage and cost, Prometheus metrics, and attempt reasons in response headers. By default Prism keeps no prompt-level logs — a few opt-in features store prompt-derived text.

The request journal stays in Prism's database for at least one year. Older records move to compressed archive files on the same machine; the archives are kept for seven years and are included in Prism's backup.

### Is there a PII guard?

Yes — opt-in. The credential and PII guard checks prompts and tool calls for emails, IPv4 addresses and phone numbers, and can observe, block or redact them. When it is on, it checks the whole request text, not only the first megabyte. The audit log never stores the matched text.

## Editions

### How do Free, Pro, Teams and Enterprise differ?

Same binary, same deployment, and each edition includes everything in the editions before it. Free costs nothing for any use, including companies, with up to 3 people, passkey sign-in and no passwords. Pro, for up to 5 people, adds model servers and GPU machines started and stopped on demand, Smart routing and semantic search in the command palette. Teams, for up to 10 people in one team, adds an organization with roles (owner, administrator, auditor and member) managed in the console, outgoing mail and export. Enterprise adds several teams, payment by invoice, several installations per subscription and support with an agreed response time, at a negotiated price from €8 per user per month before tax, with at least 5 seats. Team and person budgets set by administrators are planned. SSO and SCIM are not available.

### Can I move from Free to a paid edition later?

Yes, in place, once the paid editions go on sale: a paid edition activates in the same binary when bound to your Prism installation, keeping your connections, agent keys and history. No reinstall. A move to a higher edition takes effect at once and is charged pro rata for the rest of the period; a move to a lower edition takes effect from the next period. There is no trial: Free is how you try Prism.

### How is Prism different from other gateways?

See [how Prism compares](https://prism.tm/compare/) — side-by-side pages for LiteLLM, OpenRouter, Portkey and more, with sources.

Want to try Prism with your own providers?

[Request access](https://prism.tm/request-access/?topic=free)
