Skip to main content

Parley privacy statement

Last updated: November 12, 2026

Parley is MIT's gateway to AI models, run by MIT Information Systems and Technology (IS&T). This page explains what Parley collects when you use it, what it does with that information, who else receives it, how long it is kept, and what you can do about it. It covers every part of Parley:

  • the Parley Admin Portal (platform.parley.mit.edu), where you manage accounts, keys, knowledge bases, agents and settings
  • the Parley API, which apps, scripts and coding tools call with a Parley API key
  • the Parley CLI (parley), which sets up coding tools on your computer to use Parley
  • the Parley browser helper, the Chrome and Edge extension that collects pages behind Touchstone for a knowledge base
  • this documentation site

Parley is an MIT service, so MIT's own policies apply to it too, in particular Policy 11.1, Protection of Personal Privacy and MIT's privacy statement. This page adds the detail of how Parley itself handles your information.

Some features described here are switched on only in some Parley environments or for some accounts. Where a feature is off, nothing it would collect is collected.

The short version​

  • Parley uses your information to run the service, bill the right account and keep it secure. It does not sell it, use it for advertising or use it to train AI models, and it has no analytics or tracking.
  • What you send to an AI model goes to the company that runs that model (OpenAI, Google or Amazon Web Services) so it can answer.
  • Parley keeps a record of each request, including part of what you sent and the model's reply, for 30 days by default. An account can be set to keep no prompt or reply text at all.
  • Content you add to a knowledge base stays there until it is removed.
  • The browser helper reads only the MIT pages in a collection you start and confirm, and hands their text to Parley inside your own browser. It never reads your cookies, passwords or browsing history.

What Parley collects and why​

Who you are​

You sign in to the Portal, this site and the CLI with your MIT account, through Okta. From that sign-in and the MIT directory, Parley reads your Kerberos name, name, email address, MIT ID, affiliation (student, staff, affiliate and so on), department and, for students, year. Where an account belongs to a Moira list, Parley checks who is on it.

Parley uses this to decide what you can see and manage, to work out which accounts and monthly credit you are eligible for, to show who owns, manages or holds a key on an account, and to email you about your accounts, for example when you get a new key or an invoice. Email goes out through MIT's notification service.

Your accounts and billing​

For each account Parley keeps its name, owner, administrators, the people who hold its keys (their Kerberos names and email addresses), its budget settings, how it is billed, and the API keys issued for it.

  • Cost object billing. Each month's charges are posted to MIT's financial system as a journal voucher against the account's cost object, and can be reported to ServiceNow so they appear against an order.
  • Card billing. You type card details into a form that Stripe provides, which loads only when you add a card. The card number goes to Stripe and never reaches Parley; Parley keeps only Stripe's reference to the card. Parley gives Stripe the account's name and identifier and the billing contact's email address so Stripe can issue invoices.
  • Usage. For every request Parley records which account and key made it, the model, the number of tokens, the cost and how long it took. These records, and each account's credit history, are its billing history and do not expire.

What you send to AI models​

When you send a request (a chat message, a prompt from a coding tool, an image request, text to embed, or a file), Parley passes it to the provider that serves the model you chose and hands back the answer. Where guardrails apply, Parley's content filter (Amazon Bedrock Guardrails) screens the request and the reply first.

Parley keeps a record of each request so that a problem with it can be looked into. Besides the usage details above, the record holds the opening message of the request (often the system prompt, or your latest input) and the model's full reply; for text sent to be embedded, it holds that text. These records are kept for 30 days by default. Parley can set a different period, and if the setting cannot be read, new records are kept for one day. There is no screen or API for reading them: only Parley staff with access to its database can, when looking into a problem or a security concern.

Zero content retention. An account can be set to keep no prompt or reply text. Parley then strips the content out before the record is written and keeps only the details listed under Usage. Ask Parley support to switch it on for an account. Agent runs have to keep their input and output, so they are not available to these accounts.

When the content filter blocks something, Parley records which filters fired and how strongly, for 30 days. Parley can also choose to keep the blocked text itself, at most 4,000 characters, to tune the filter. That is off unless Parley staff turn it on, and it never applies to an account with zero content retention.

Files you attach through the API's files endpoint go straight to the provider (OpenAI or Google). Parley keeps only the file's name, type and size, and has the provider delete the file about two days after upload.

Knowledge bases​

A knowledge base holds what an account adds to it: files people upload, and pages its connectors fetch (a website crawl, an API, an SFTP folder, or pages behind Touchstone collected with the browser helper or the CLI). Parley stores this content in its own storage on Amazon Web Services and indexes it with Amazon Bedrock so it can be searched. Each item records who added or collected it. Usernames, passwords and tokens a connector needs are kept in AWS Secrets Manager and never shown again.

Content stays until someone who manages the knowledge base removes it. Removing a file takes it out of search straight away. Parley's storage also keeps earlier versions of files, as a safeguard against accidental loss, so a copy can remain there after the file is removed.

When a knowledge base is searched, Parley records the question, which knowledge bases were searched and references to what was found, but not the passages themselves. That search history is kept for 30 days by default. Account owners and administrators can see the questions; Parley staff see the references and other details but not the question itself. For an account with zero content retention, the question is not recorded. The passages a search finds go to the AI model along with the rest of the request.

Tools, connected services and agents​

Where they are available, Parley lets AI models use tools for you: MIT systems such as the Data Warehouse, Roles and People Search, your Outlook mail and calendar, web search, and other MCP servers that Parley or your account has added. A tool acts with your own access, never anyone else's.

  • Saved secrets. Secrets you save on the Portal's Secrets page, such as a Data Warehouse login, are stored in AWS Secrets Manager under your Kerberos name. Only you can see or change them, from your own Portal sign-in, and they are never logged. Deleting one deletes it at once.
  • Sign-in connections. When you connect Outlook or another service, its sign-in token is held in the token vault of Amazon Bedrock AgentCore, not by Parley. Parley keeps only whether you are connected, and you can disconnect at any time.
  • Tool calls. Every tool call goes into Parley's audit trail: who made it, which tool, from where, whether it worked and how long it took, plus only the arguments that tool is set up to record. Results are never recorded.
  • Web search sends your search query to Perplexity, which runs the search.
  • Agent runs keep their input, instructions, output and progress for 90 days so the result can be read. Through the API, only the key that started a run can read it; in the Portal, the account owner can see its input and output too.

Collected automatically​

  • Where requests come from. For each API key, Parley notes the IP address, the client's user agent, the kind of client and whether the request came from inside MIT's network. This is how Parley spots a key being used from somewhere unexpected. Each entry is kept for 180 days after it was last seen.
  • Audit trail. Changes made through the Portal and the management parts of the API (accounts, keys, settings, secrets) are recorded with who made them, when, from which IP address and user agent, and what was sent, with secret values removed. The audit trail is kept for 400 days.
  • Service logs. Parley's servers keep operational logs, such as errors and timings, for 18 months. They leave out prompts, replies and credentials, but can say who made a request, for example by Kerberos name or email address. The load balancer in front of the API keeps standard access logs (IP address, time and address requested) for 90 days.
  • The CLI sends its version, your operating system and a random install identifier with its calls to Parley, so Parley can count installs and see when they update. The identifier is random and is not worked out from anything on your computer.

The websites​

The Portal and this site have no analytics, advertising or tracking cookies. Your sign-in tokens and display preferences (theme, date ranges, filters, which notices you have read) are kept in your browser's local storage. The Portal loads its fonts from Google Fonts, and Stripe's card form only when you add a card, so Google and Stripe see your IP address when your browser fetches them.

The Parley browser helper​

The browser helper is an extension for Chrome and Edge. Its one job is to collect the text of MIT pages behind Touchstone, in your own signed-in browser, for a Parley knowledge base. It does nothing until you start a collection from the Portal.

  • What it reads. When you start a collection, Parley works out which pages it covers and the helper shows you the list. Once you confirm, it opens each page in an ordinary tab and takes the page's title and main text, or the links on a listing page. It also looks at the start of each page to tell whether Touchstone is asking you to sign in; that check stays inside the helper.
  • Which sites. Only MIT sites (*.mit.edu), and only the ones you allow. The helper has no access to any site when it is installed. When you press Collect, Chrome asks you to allow the sites in that collection.
  • What it never reads. Your cookies, passwords, sign-in tokens or browsing history, or anything outside mit.edu. It never sends your Touchstone session anywhere.
  • Where the text goes. To the Portal page that started the collection, in the same browser. That page adds it to the knowledge base with your own Parley sign-in, and each page records that you collected it. From there it is handled like any other knowledge base content (see Knowledge bases). The helper makes no network requests of its own and holds no Parley credentials; the pages it opens load in your browser just as they would if you visited them.
  • On Parley's own pages. A small script runs on Portal pages only, to pass messages between the page and the helper. It reads nothing on those pages.
  • What it stores. While a collection runs, the helper keeps its list of pages in the browser's session storage, which lives in memory and is cleared when the helper's window closes. Nothing is written to disk.
  • Pages that show your own details. The helper reads a page as you see it. If a site shows personal details to whoever is signed in, those would be collected too, so collect only pages that look the same for everyone.
  • Taking access back. Open the helper from the browser toolbar and choose Remove next to a site, or remove the helper altogether in chrome://extensions.

The Parley CLI​

The CLI runs on your computer, and most of what it handles stays there.

  • Your API key is kept in your computer's password store, or in a file only you can read where there is no password store. Your Okta sign-in tokens are kept in the CLI's settings folder, also readable only by you.
  • It writes settings for Claude Code, Codex and OpenCode so they use Parley, and removes other keys from their settings that would get in the way.
  • parley ps reads those tools' session files to list your recent sessions. They are not sent anywhere.
  • To check for updates it downloads a version list from Parley's public download site, and it can download Node.js from nodejs.org when a tool needs it.
  • parley kb harvest uses a temporary browser profile for the collection and deletes it afterwards.

Who else receives information​

Parley shares information only with the services it needs in order to work, and only what each one needs.

WhoWhat they receiveWhy
OpenAI, Google, and Amazon Web Services through Amazon Bedrock (which also serves Anthropic's Claude models)The requests you send to their models, and files you attachTo answer your request
Amazon Web ServicesEverything Parley storesParley runs on AWS
StripeThe account's name and identifier, the billing email address, and the card details you enter in Stripe's formCard payments and invoices
PerplexityYour search queryWeb search, where it is available
MicrosoftRequests to your Outlook mail and calendar, within the access you grantedOutlook tools, once you connect them
MIT systems (directory, Moira, Finance, ServiceNow, Data Warehouse, Roles, People Search, notifications)What each one needs, such as your Kerberos name, an account's cost object and charges, or a query you runRunning and billing the service

Each AI provider handles what it receives under its own terms for use of its API. Parley does not sell information, does not use it for advertising, and does not share it with anyone else except where the law or MIT policy requires.

Where it is kept​

Parley stores information in Amazon Web Services in the United States, in Northern Virginia with a copy in Northern California for resilience. AI providers may process a request elsewhere. Some Claude models on Amazon Bedrock are served through Amazon's global routing, so those requests can be processed outside the United States.

Who at MIT can see it​

  • People on your accounts see what their role on the account allows. Owners and account administrators manage the account and see its usage, keys and knowledge base search questions. People who only hold a key see less.
  • Parley staff, the IS&T team that runs Parley, can see account, usage and billing information in order to support, bill and secure the service, and can read stored request records when looking into a problem. They work under MIT's policies on personal privacy.
  • MIT's finance and service management teams see the charges posted to cost objects.

How long it is kept​

InformationKept for
Request records, with prompt and reply text30 days by default; no text at all with zero content retention
Usage, charges and credit historyNo expiry, as billing records
Knowledge base contentUntil removed; earlier versions of files can remain in storage
Knowledge base search history30 days by default
Agent runs90 days
Content filter blocks30 days
Files attached through the APIAbout two days
Where API requests come from (IP address, client)180 days after last seen
Audit trail400 days
Service logs18 months
Load balancer logs90 days
CLI install recordsNo expiry
Saved secrets and sign-in connectionsUntil you delete or disconnect them
Accounts, keys and the people on themFor the life of the account, including after it is deactivated

Your choices​

  • Ask Parley support to switch on zero content retention for an account.
  • Remove files and connectors from a knowledge base.
  • Delete saved secrets and disconnect services on the Portal's Secrets page.
  • Pause, regenerate or delete API keys in the Portal.
  • Take back the browser helper's access to a site, or remove the helper.
  • Ask what Parley holds about you, or ask for something to be corrected or deleted, by emailing parley-support@mit.edu. Billing records may need to be kept for MIT's financial records.

Security​

Parley is reached over HTTPS only, and keeps its data in AWS services that encrypt it at rest. Secrets and connector credentials are held in AWS Secrets Manager.

Changes and questions​

When Parley changes how it handles information, this page is updated and the date at the top changes. Questions about this page or about your information go to parley-support@mit.edu.