# Actions
Source: https://docs.anagram.ai/actions/overview
Things your agent can do beyond answering questions.
Actions give your agent hands. Answering questions is the baseline; actions let it show a store locator, push a lead into Klaviyo, or log survey answers to a spreadsheet, right in the middle of a conversation.
You manage them on the **Actions** page. Each action has an on/off toggle and its own configuration. Actions are set up once for your whole agent and work across every placement.
## Show stores
Displays an interactive map and list so shoppers can find your physical locations. Click **Set up stores**, add your locations, and enable the action. Stores save as you add them.
Once enabled, the agent can bring up the locator whenever a shopper asks where to buy in person. You can also make it the very first thing shoppers see on an Embedded inline placement, using the **Show stores** [suggestion mode](/launch/conversation-starters#show-stores).
## Send to Klaviyo
Lets the agent send a shopper's details to a Klaviyo list, so conversations can feed your email and SMS flows. Someone asks about a sold-out item, the agent offers to notify them, and their profile lands in Klaviyo.
This action needs your Klaviyo account connected first; the card walks you through it. After configuring, use **Test connection** to verify the list is reachable, or **Send a sample profile** to push a real test profile into the list and see it land.
## Send to Google Sheet
Appends a row to a Google Sheet you choose. Handy for collecting structured answers without more tooling: wholesale inquiries, feedback, quiz responses.
Connect your Google account, pick the spreadsheet, then use **Test connection** to confirm Anagram can see it, or **Send a sample row** to write a real test row.
The two sample tests write real data: a real profile in Klaviyo, a real row in your sheet. Delete the test entries afterward if they'd get in the way.
## Telling the agent when to act
Enabling an action makes it available; the agent decides when to use it based on the conversation. If you want more control over the moment, write a [skill](/brain/skills). For example, a skill triggered by "when a shopper asks to be notified about restocks" can instruct the agent to collect an email and send it to Klaviyo, in that order, every time.
# Analytics
Source: https://docs.anagram.ai/analytics/overview
What each card on the Analytics page means, and exactly how the numbers are counted.
The **Analytics** page answers the question your CFO will ask: is this thing working? Use the date range picker at the top to set the reporting window (it defaults to the last 30 days). This page walks through every card in the order it appears, but start with the three words that trip everyone up.
## Messages, engagements, and visitors
These are three different counts, and they nest inside each other:
* A **message** is one thing a shopper sends: a typed question or a tapped suggestion chip. A single conversation usually has several.
* An **engagement** is a chat. Specifically, it's a conversation in which the shopper sent at least one message. Ten messages back and forth in one chat is still one engagement. A shopper who opens the widget, looks, and leaves without sending anything is zero engagements; seeing the agent doesn't count.
* An **engaged visitor** is a person. If the same shopper chats with your agent on Monday and again on Thursday, that's two engagements but one engaged visitor.
A worked example: one shopper has two conversations, sending four messages in the first and two in the second. That's 6 messages, 2 engagements, 1 engaged visitor.
Most of the page counts engagements, because a chat is the natural unit of "the agent did some work." Engaged visitors show up where people are the right unit, like conversion rate.
Your own preview chats inside Anagram never count. Every number on this page comes from real shoppers on your live site.
## The cards, top to bottom
### Usage
Total engagements in the period, with a chart of engagements per day. This is your "are shoppers using it" number, and the suggestion chips on your [placements](/launch/conversation-starters) are usually the biggest lever on it. Click **View details** to open the individual conversations behind any day.
### Anagram-assisted sales
The number of orders attributed to the agent in the period, with a per-day chart. An order counts as Anagram-assisted when the shopper interacted with your agent and then completed the purchase within the attribution window, currently 7 days after the interaction.
A few details worth knowing about how this is counted:
* Attribution is anonymous. Anagram recognizes the same browser across the chat and the checkout; nobody has to log in or click a special link.
* It is not last-click. The agent doesn't need to be the final thing the shopper touched before buying. If they chatted, left, came back through an email two days later and purchased, that order still counts, because the chat happened within the window.
* Each order counts once, even if the shopper had several conversations.
* It requires your [Shopify connection](/shopify), which is how orders get linked back to conversations.
Because the window looks forward from the chat, a sale can land a few days after the engagement that earned it. Give a new launch at least a week before judging this number.
### Attributed revenue
The money side of the same thing: total order value of your Anagram-assisted sales for the period, shown in your store's currency. Same attribution rules as above; this is the sum where assisted sales is the count.
### Conversion Rate
The share of engaged visitors who went on to purchase within the attribution window. This one is counted in people, not chats: purchased visitors divided by engaged visitors. Hover over the number and the card shows you the exact breakdown it was calculated from.
Counting people rather than chats keeps the rate honest. A shopper with three conversations and one purchase is one converted visitor out of one engaged visitor, not one out of three.
### Average Order Value
Attributed revenue divided by the number of assisted orders: what the typical agent-assisted order is worth. Comparing this against your store-wide AOV is one of the more persuasive numbers here, since guided shoppers often buy more. The card shows a dash until you have at least one assisted sale.
### Top products
Which products sell after agent conversations, with units sold and assisted revenue per product. Good for spotting what the agent is effective at selling, and occasionally what it recommends that people don't buy.
### Conversations
Every transcript, searchable, with a **Purchases** filter to read only the conversations that ended in an order. Each conversation shows which placement it came from.
This is the most underrated section on the page. Read ten transcripts and you'll find missing knowledge, an unclear policy page, or a question you never thought shoppers would ask. Fixes usually take a minute in [Brain](/brain/knowledge).
## How the date range interacts with attribution
The date range selects when the engagements happened. Assisted sales are then allowed to complete inside the attribution window even if that runs past the end of the range. So "last week" means orders attributed to last week's conversations, including a purchase that closed this Monday. This also means recent periods can tick upward for a few days as their window plays out.
## A note on scope
Analytics covers your whole agent across all placements. Placement shows up as a label on each conversation so you can see where chats happen, but the numbers are agent-wide.
# Knowledge
Source: https://docs.anagram.ai/brain/knowledge
The facts your agent draws on: your catalog, your website, and anything you add yourself.
Knowledge is the first tab in **Brain**, and it's where most of your agent's answers come from. There are three sources, and they work together. When a shopper asks something, the agent pulls whatever is relevant, whether that's a product from your catalog, a page from your site, or a fact you typed in.
## Catalog
Your product catalog is the backbone. Connect Shopify and Anagram syncs your products automatically every day, including titles, descriptions, prices, tags, collections, and metafields. The agent uses this to recommend products, answer questions like "does this come in a medium?", and show product cards in chat.
To connect, open **Brain**, find the Catalog card, and click **Connect Shopify**. You'll enter your `myshopify.com` domain and approve the app install. Full details are on the [Shopify](/shopify) page.
Once connected, click **Manage** on the Catalog card to browse and search everything the agent can see. If a product looks wrong here, fix it in Shopify and it will update on the next sync.
## Web search
With web search on, the agent can look up answers on your website in real time. This is how it handles questions your catalog can't: shipping policies, return windows, brand story, ingredient pages.
The **Web search** card has a few settings:
* **Site to search** limits the agent to a specific domain. Leave it blank and it uses your store's domain.
* **Also search the open web** lets the agent look beyond your site. Most brands leave this off so answers stay grounded in their own content.
Web search is on by default. If your site's policy pages are accurate and current, you're in good shape without any extra work.
## Added knowledge
This is for the facts that live in your head or your team's Slack, not on your website. The card calls them tribal facts, and that's the right way to think about it. Examples:
* "Our candles are made from soy wax"
* "We don't offer gift wrapping during December"
* "The Redwood line is being discontinued; don't recommend it for restocks"
Click **New entry**, give it a **Title**, write the fact, and click **Add entry**. Each entry is a single fact or a small cluster of related ones. Short, specific entries work better than long documents because the agent retrieves entries individually when they're relevant.
When you test your agent and it gets something wrong, the fix is usually one knowledge entry away. Ask yourself: where would the agent have found the right answer? If the answer is "nowhere," add it here.
## When changes take effect
Saving a knowledge entry makes it live for shoppers within a few minutes. Catalog changes flow in with the daily Shopify sync, or sooner if you trigger a manual sync.
# Rules
Source: https://docs.anagram.ai/brain/rules
Always-on instructions that keep your agent on-brand and out of trouble.
Rules are standing orders. Every rule you save applies to every response your agent gives, on every page, in every conversation. They're how you encode the things a good employee would just know: what to promise, what never to promise, how to talk about competitors, when to point someone to customer service.
You'll find Rules as a tab in **Brain**.
## Writing a rule
Each rule is one instruction with a polarity:
* **Always** rules describe what the agent should do. "Always greet shoppers warmly and stay on-brand."
* **Never** rules describe what it must avoid. "Never promise a specific delivery date."
Pick **Always** or **Never**, write the behavior in the text box, and click **Save rule**. It's live immediately. Your saved rules appear in a list below the composer, and every rule in that list is active. To retire a rule, delete it; there's no on/off switch.
## What makes a good rule
Keep each rule to a single, concrete behavior. A few patterns that work well:
* **Guardrails.** "Never give medical advice; suggest the customer consult a doctor."
* **Brand voice.** "Always refer to our loyalty program as The Circle, not the rewards program."
* **Sales behavior.** "Always suggest a complementary product after answering a product question."
* **Escalation.** "Never handle order cancellations; direct shoppers to [support@yourbrand.com](mailto:support@yourbrand.com)."
Avoid stuffing several instructions into one rule. "Always be friendly, mention free shipping, and ask for their email" is three rules pretending to be one, and the agent follows three separate rules more reliably.
## Rules vs. knowledge vs. skills
It's worth keeping the three straight, because the fix for a bad answer depends on which one is missing:
* The agent got a **fact** wrong or didn't know it? Add [knowledge](/brain/knowledge).
* The agent **behaved** wrong, was off-brand, or promised something it shouldn't? Add a rule.
* The agent needs to follow a **multi-step process** for a specific task, like size finding? Write a [skill](/brain/skills).
## Testing a rule before saving
While you're writing a rule, the preview chat next to the editor automatically applies it on top of what you've already saved. You'll see a "Testing unsaved changes" badge; nothing there goes live or touches your analytics. Chat with the preview, check the behavior change, and save once it looks right. This works for edits to knowledge and skills too.
# Skills
Source: https://docs.anagram.ai/brain/skills
Step-by-step playbooks your agent follows for specific shopper tasks.
A skill is a playbook. Where [rules](/brain/rules) shape every response, a skill only kicks in when a shopper's situation calls for it, and then the agent follows your instructions step by step. Think of the difference between "always be polite" (a rule) and "here's exactly how we help someone find their size" (a skill).
You'll find Skills as a tab in **Brain**.
## Creating a skill
Click into the Skills tab and fill out three fields:
**Skill name.** A short label like "Size finder" or "Gift concierge." This is for you and the agent, not shoppers.
**When should this skill be used?** This is the trigger, and it's the most important field. It controls when the agent reaches for this skill, so be specific about the shopper intent and the signals that should activate it. For example: "When a shopper asks which size, fit, or variant will work for them."
**Instructions.** The playbook itself. Write it as numbered steps in plain language, the way you'd train a new hire:
```text theme={null}
1. Ask what they usually wear in other brands.
2. Ask whether they prefer a snug or relaxed fit.
3. Recommend a size using our fit guide: our tops run small,
so suggest sizing up for a relaxed fit.
4. Offer to show the product in their size.
```
All three fields are required. Save, and the skill is live. Like rules, every skill in your list is active; delete a skill to retire it.
## How the agent uses skills
The agent always knows which skills exist and what triggers them, but it only reads the full instructions when a conversation matches a trigger. So a shopper asking about shipping never sees size-finder behavior, and your size-finder instructions can be as detailed as you want without cluttering every other conversation.
## Skill ideas for a Shopify brand
* **Size and fit finder.** Ask fit questions, apply your sizing quirks, recommend a size.
* **Gift concierge.** Ask who the gift is for, the occasion, and budget, then recommend two or three options.
* **Routine builder.** For skincare and supplements, ask about goals and build a multi-product routine.
* **Restock helper.** When someone mentions running out, find their product and offer the larger size or a bundle.
## Skills vs. Actions
Skills are conversational playbooks: instructions the agent follows in chat. [Actions](/actions/overview) are integrations: concrete things the agent can do, like showing a store locator map or sending a lead to Klaviyo. They combine well. A "help them find a store" skill can tell the agent when and how to use the Show stores action.
# How Anagram works
Source: https://docs.anagram.ai/concepts/site-agent
One agent, one brain, placed wherever you want it on your site.
## One agent per store
Your Anagram project has exactly one Site Agent. It has one brain, one voice, and one look, no matter how many places it appears on your site. That's deliberate. Shoppers who chat with the bar on your homepage and later open the button on a product page are talking to the same assistant, with the same knowledge and the same rules.
This is different from tools where you build a separate bot per page. In Anagram you configure the agent once, then decide where it shows up.
## The pieces
**Brain** is everything the agent knows and how it behaves. It has three parts:
* [Knowledge](/brain/knowledge) is the facts: your product catalog, your website, and anything you type in yourself.
* [Rules](/brain/rules) are always-on instructions, phrased as things the agent should always or never do.
* [Skills](/brain/skills) are step-by-step playbooks the agent follows for specific tasks, like walking someone through finding their size.
**Placements** are where the agent appears. Each placement has a form factor (center bar, floating button, custom trigger, or an embedded widget) and its own [conversation starter](/launch/conversation-starters): the greeting and suggested questions shoppers see first. A custom trigger lets a button you already own open an Anagram panel. Placements live under **Launch**.
**Style** controls how the agent looks: colors, fonts, and branding. It's set once for the whole agent, on the **Style** tab under Launch. Individual placements can override parts of it if you need to.
**Actions** are things the agent can do beyond answering: show a store locator map, send a shopper's details to Klaviyo, or log responses to a Google Sheet. See [Actions](/actions/overview).
**Analytics** shows engagements, Anagram-assisted sales, and every conversation transcript. See [Analytics](/analytics/overview).
## Saving goes live
There are no drafts. When you save a rule, a knowledge entry, or a placement change, it's live for shoppers within a few minutes. The way to try something before committing is the preview chat, which lets you test unsaved edits against the real agent. Save when it behaves the way you want.
## What shoppers experience
A shopper lands on your site and sees the agent in whatever form you placed there, with a greeting and a few tappable suggested questions. They tap one or type their own. The agent answers using your catalog and knowledge, shows product cards when it recommends something, and follows your rules the whole way. If they buy within the attribution window after chatting, that sale shows up in your analytics as Anagram-assisted.
# Build a recommendation quiz
Source: https://docs.anagram.ai/guides/recommendation-quiz
Turn your agent into a guided product finder with one skill and one suggestion.
Product quizzes convert because they do the shopper's thinking for them. The usual version is a rigid form built in a quiz app. Your agent can run the same play conversationally: ask a few questions, then recommend real products from your catalog, and handle the follow-up questions a form never could.
This guide builds one with a single [skill](/brain/skills). Plan on 15 minutes.
## Before you start
Your [catalog](/brain/knowledge#catalog) needs to be connected, since the recommendations come from your synced products. That's it.
## Step 1: Write the quiz as a skill
Go to **Brain → Skills** and create a new skill. Here's a worked example for a skincare brand; adapt the questions and logic to your products.
**Skill name**
```text theme={null}
Routine finder quiz
```
**When should this skill be used?**
```text theme={null}
When a shopper asks for a recommendation, says they don't know
what to choose, asks to take the quiz, or asks what's right for
their skin.
```
The trigger is what makes the quiz fire at the right moments, so cover the different ways shoppers express "help me choose." Notice it isn't just "when they ask for the quiz"; anyone signaling indecision is a quiz candidate.
**Instructions**
```text theme={null}
Run a short quiz to find the right routine. Ask one question at a
time and wait for the answer before asking the next.
1. Ask about their skin type: oily, dry, combination, or not sure.
If they're not sure, ask whether their skin feels tight after
washing (dry) or gets shiny by midday (oily).
2. Ask what their main goal is: clearer skin, fewer fine lines,
more hydration, or evening out tone.
3. Ask if they have any sensitivities or ingredients they avoid.
4. Recommend 2 or 3 products from the catalog that fit their
answers, as a routine: cleanser first, then treatment, then
moisturizer. For each one, say in one sentence why it fits
what they told you.
5. Ask if they'd like alternatives at a different price point.
Keep it light. Never diagnose skin conditions; if they mention a
medical concern, suggest they see a dermatologist.
```
A few things this example does that are worth copying:
* **One question at a time.** Without that line, the agent may dump all the questions at once, which kills the quiz feel.
* **A recovery path for "not sure".** Shoppers who'd bounce off a form dropdown will answer "does your skin feel tight after washing?"
* **A capped recommendation count.** Two or three products with reasons beats a wall of options.
* **A closing offer.** The price-point question keeps the conversation going instead of ending on a list.
Save the skill.
## Step 2: Test it
Use the preview chat and come at the quiz from different angles: "what should I buy?", "I'm new to skincare", "take the quiz". Check that it triggers each time, asks one question at a time, and recommends sensible products.
If it recommends oddly, the fix is usually in the instructions: add a line like "never recommend the exfoliating toner for sensitive skin." If a product detail is wrong, that's a catalog or [knowledge](/brain/knowledge) fix instead.
## Step 3: Make the quiz discoverable
The skill only helps shoppers who talk to the agent, so put the quiz in front of them:
* **Add a suggestion chip.** In your placement's [suggestions](/launch/conversation-starters), switch to Fixed suggestions and add "Find my routine" or "Take the 30-second quiz" as a chip, or keep Auto mode and rely on the agent surfacing it. A fixed chip is the more reliable pull.
* **Pin it where it's relevant** with a [suggestion set](/launch/suggestion-sets). A set targeting your bestselling collection can open with "Not sure which is right for you? Take the quiz."
* **Give the quiz its own placement.** An [Embedded inline](/launch/placements) placement on a "Find your routine" landing page turns the quiz into a destination you can link from email and ads.
## Variations
The same skeleton works for most guided-selling problems. Change the questions and the recommendation logic:
* **Gift finder**: who's it for, occasion, budget, then two or three options.
* **Fit finder**: current brand and size, fit preference, then one size recommendation with your brand's sizing quirks baked into the instructions.
* **Bundle builder**: their goal, what they already own, then a set of products that work together.
If you build one you're proud of, we'd genuinely like to see the transcript. Send it to [support@anagram.ai](mailto:support@anagram.ai).
# Anagram MCP
Source: https://docs.anagram.ai/integrations/mcp
Ask ChatGPT, Claude, Codex, or Cowork questions about your Anagram conversations and analytics.
The Anagram MCP connects your AI assistant to Anagram Studio. You can ask questions about shopper conversations, conversation activity, commerce impact, and other analytics without exporting data first.
The connection is read-only. It cannot change your Anagram projects or settings, and it only returns data from projects your Anagram account can access.
## Connection details
Use these values in every client:
| Field | Value |
| -------------- | ----------------------------------- |
| Name | `Anagram` |
| MCP server URL | `https://studio.anagram.ai/api/mcp` |
| Authentication | OAuth |
During setup, your browser opens an Anagram sign-in page. Sign in and approve the connection. Anagram uses Clerk to complete the OAuth flow. You do not need to create an OAuth client ID or secret.
## Choose your client
Install the official Anagram plugin from ChatGPT's Plugins Directory.
Add Anagram as a custom connector in Claude.
Connect from the ChatGPT desktop app, Codex CLI, or Codex IDE extension.
Add Anagram to your Claude account, then use it in Cowork.
## Try the connection
After connecting, start with a question such as:
* "Which Anagram projects can I access?"
* "What did shoppers ask about most often this week?"
* "Show me the products with the most assisted sales last month."
* "Find conversations where shoppers asked about sizing."
Analytics pass through a buffered event pipeline, so the newest activity may take a short time to appear.
# Connect Anagram to ChatGPT
Source: https://docs.anagram.ai/integrations/mcp-chatgpt
Install the official Anagram plugin from ChatGPT's Plugins Directory.
Anagram is available as an official plugin in ChatGPT's Plugins Directory. You do not need developer mode or an MCP server URL.
## Connect Anagram
Open **Plugins** from the ChatGPT sidebar, or go to **Settings → Plugins**. Search for `Anagram` and open its listing.
Select the plus button to install Anagram. ChatGPT will prompt you to connect your Anagram account.
Select **Connect**, sign in to Anagram, and approve access. ChatGPT completes the connection securely through OAuth.
Start a new chat and ask ChatGPT to use Anagram. For example: "Use Anagram to show me what shoppers asked about most often this week."
If **Connect** is unavailable or shows **Disabled by admin**, ask your ChatGPT workspace admin to enable Anagram. Plugin availability can depend on your plan, workspace settings, role, and region.
The Anagram connection is read-only and only returns data your Anagram account can access.
See [OpenAI's Plugins Directory guide](https://help.openai.com/en/articles/20001256-plugins-in-chatgpt-and-codex) for plan and workspace details.
# Connect Anagram to Claude
Source: https://docs.anagram.ai/integrations/mcp-claude
Add the Anagram MCP as a custom connector in Claude.
Anthropic calls remote MCP connections "custom connectors." The connection is saved to your Claude account, so you can use it in Claude on the web, desktop, and mobile.
## Individual accounts
In Claude, open **Customize → Connectors**.
Select **+ → Add custom connector**.
Enter:
* **Name:** `Anagram`
* **Remote MCP server URL:** `https://studio.anagram.ai/api/mcp`
Leave the OAuth Client ID and OAuth Client Secret in **Advanced settings** empty. Select **Add**.
After selecting **Add**, complete the Anagram sign-in and approval screen when Claude asks. OAuth is handled automatically through Clerk.
Start a conversation. Select **+ → Connectors**, then turn on **Anagram**.
Free Claude accounts can add one custom connector.
## Team and Enterprise accounts
An Owner or Primary Owner must add Anagram to the organization first:
1. Open **Organization settings → Connectors**.
2. Select **Add**, hover over **Custom**, then select **Web**.
3. Add `https://studio.anagram.ai/api/mcp`. Leave the optional OAuth client ID and secret empty.
4. Select **Add**.
Each member can then open **Customize → Connectors**, find **Anagram** with the **Custom** label, and select **Connect** to sign in.
Adding Anagram to an organization does not give every member access to the same data. Each person signs in separately, and Anagram only returns projects that person can access.
See [Anthropic's current custom connector guide](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) for plan and workspace details.
# Connect Anagram to Codex
Source: https://docs.anagram.ai/integrations/mcp-codex
Add the Anagram MCP to the Codex desktop, CLI, and IDE clients.
Codex shares MCP settings across the ChatGPT desktop app, Codex CLI, and Codex IDE extension.
## ChatGPT desktop app
In the ChatGPT desktop app, open **Settings → MCP servers** and select **Add server**.
Set the name to `Anagram`, choose **Streamable HTTP**, and enter:
`https://studio.anagram.ai/api/mcp`
Save the server, then select **Restart**.
Find Anagram in the server list and select **Authenticate**. Complete the Anagram sign-in and approval screen. OAuth is handled automatically through Clerk.
Type `/mcp` in the Codex composer. Anagram and its tools should appear in the connected server list.
## Codex CLI
Run:
```bash theme={null}
codex mcp add Anagram --url https://studio.anagram.ai/api/mcp
codex mcp login Anagram
```
Your browser opens the Anagram sign-in and approval screen. After signing in, run `codex mcp list` to check the connection.
## Configuration file
You can also add Anagram to `~/.codex/config.toml`:
```toml theme={null}
[mcp_servers.Anagram]
url = "https://studio.anagram.ai/api/mcp"
auth = "oauth"
```
Then run:
```bash theme={null}
codex mcp login Anagram
```
The IDE extension reads the same configuration. A trusted project can instead keep the entry in `.codex/config.toml`.
## Codex IDE extension
The IDE extension does not have a separate add-server form. Open its settings and choose the option to edit `config.toml`, add the configuration above, then run `codex mcp login Anagram` in a terminal.
See [OpenAI's current Codex MCP guide](https://developers.openai.com/codex/extend/mcp) for all configuration options.
# Connect Anagram to Cowork
Source: https://docs.anagram.ai/integrations/mcp-cowork
Add Anagram to your Claude account and use it in Cowork.
Cowork uses the remote connectors saved to your Claude account. You do not create a second connection inside Cowork.
## Connect Anagram
In Claude, open **Customize → Connectors**.
Select **+ → Add custom connector**, then enter:
* **Name:** `Anagram`
* **Remote MCP server URL:** `https://studio.anagram.ai/api/mcp`
Leave the OAuth Client ID and OAuth Client Secret in **Advanced settings** empty. Select **Add**.
After selecting **Add**, complete the Anagram sign-in and approval screen when Claude asks. OAuth is handled automatically through Clerk.
Start a Cowork session by selecting **Cowork** in the message box on the Claude home screen. In the session, select **+ → Connectors**, then turn on **Anagram**.
If you use a Team or Enterprise account, an Owner or Primary Owner must add the custom connector under **Organization settings → Connectors** before members can connect it.
Anagram is a remote connector. It works through your Claude account and does not require the Claude Desktop app to stay open.
Cowork on web and mobile is in beta and rolling out by plan. If you do not see **Cowork** in the message box, update the app and check Anthropic's current availability.
See [Anthropic's current connector guide](https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities) and [Cowork client guide](https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile) for current availability.
# Introduction
Source: https://docs.anagram.ai/introduction
Anagram puts an AI shopping agent on your store that answers questions, recommends products, and helps shoppers buy.
## Meet your Site Agent
Anagram gives your store a **Site Agent**: an AI assistant that lives on your site and talks to shoppers the way your best salesperson would. It answers product questions, recommends items from your catalog, helps people find the right size or a nearby store, and nudges browsers toward checkout.
You set it up once. There is one agent per store, and it shows up wherever you place it: as a search-style bar in the middle of the page, a floating chat button, or a panel embedded right into a page.
## How it fits together
Everything you do in Anagram happens in four areas:
What your agent knows and how it behaves. Knowledge, Rules, and Skills live here.
Where your agent appears on your site and what shoppers see first.
Things your agent can do beyond chatting, like showing store locations or sending leads to Klaviyo.
Engagements, assisted sales, and full conversation transcripts.
## Where to start
If you're new, the [onboarding](/onboarding) page explains the setup wizard you'll see on first sign-in, and the [quickstart](/quickstart) walks through the four-step launch checklist on your Anagram home page: add knowledge, test the agent, configure a placement, and put the snippet on your site.
If something isn't working, check [troubleshooting](/troubleshooting) first. If you're still stuck or can't find what you need in these docs, email [support@anagram.ai](mailto:support@anagram.ai). A real person reads it.
# Add to your site
Source: https://docs.anagram.ai/launch/add-to-your-site
Put the code snippet on your storefront and go live.
Once a placement is configured, getting it onto your site is one small code snippet. Click **Launch** on your Anagram home page, or **Copy snippet** on the placement, and you'll get something like this (copying the snippet is where your [plan or free trial](/plans-and-billing) kicks in):
```html theme={null}
```
Each placement has its own snippet. The snippet never changes after you paste it; everything you edit later in Anagram, from suggestions to style to the form factor itself, flows through automatically. The one exception worth knowing: if you switch a placement between a floating form factor and an embedded one, you may want to move the snippet, since embedded widgets render exactly where the snippet sits.
## Where to paste it
It depends on the form factor:
* **Center bar** and **Floating button** float above the page, so the snippet can go almost anywhere. The convention is just before the closing `