Chat with the agent
This page covers the chat itself: where to open it, the message box and its options, what a reply shows, and how to manage your chats. The AI Agent is available on Enterprise plans. For who can use it, see AI Agent.
Open the agent
Click Agent in the left navigation. It is marked "beta". The Agent page opens with your chats on the left and a new chat on the right.
If you don't see Agent, the agent isn't on for you in this workspace. See Who can use it. If you open the Agent page while agents are off, you see "Agents are off in this workspace":
| You are | What the page says |
|---|---|
| A Workspace Owner | "You turned agents off for everyone in this workspace. Your settings and the access you gave are kept." Click Open AI agent settings to turn them on. |
| Anyone else | "The workspace owner has turned agents off, so there is no agent here for now. Everything else works as before." |
Dots on the Agent icon
A dot on the Agent icon tells you a chat has news, so you don't have to keep the page open.
| Dot | Meaning |
|---|---|
| Amber | A chat is waiting on you: a plan to approve, a question, a failed reply, or work stopped at its limit |
| Blue | A chat has a new result |
Hover the icon to see counts, for example "2 chats need you · 1 new result". Click it while it's amber to open the chat that waits on you.
When long work (more than a minute) finishes while Mammoth is in a background tab, the tab's title shows a count, for example "(1) Mammoth", and its icon gets a dot until you come back.
Finished work, plans waiting for you and alerts also appear in the Agent group of the notifications bell. See Agent notifications.
Ask about a dataset or your automations
Some pages have their own chat, docked beside the page and about that one thing. Click Ask at the right of the page's top bar, or press ⌘J (Ctrl+J on Windows).
| Where | Ask opens | Suggestions |
|---|---|---|
| A dataset's page | "Ask about this dataset": its source and refresh, its columns, or what a column means | "Where does this data come from?", "When was it last refreshed?", "What does each column mean?" |
| Data Library > Automations tab | "Ask about these automations": every automation in the project | "Which automations failed this week?", "What runs every morning?", "Which ones are paused?" |
| The same tab, with one automation selected | "Ask about this automation": when it runs, who gets it, or why it didn't fire. A change you ask for waits for your OK. | "When does this run next?", "Why didn't it fire last time?", "Who gets it?" |
Click a suggestion to send it, or type your own question in the box ("Ask about this dataset…").
- Each dataset or automation has one chat per person. Come back later and your earlier messages are still there.
- The chat is 360 pixels wide and sits beside the page. In a window narrower than 1,100 pixels it lies over the page instead.
- Close it with the ✕ in its header, Esc, or ⌘J again. Mammoth remembers whether you left it open, and opens it the same way on the next dataset.
- A dot on Ask means this chat has news: blue for a new result, amber when it waits on you.
From a main chat
If you open a dataset from a chat on the Agent page, the docked chat shows that chat's messages about the dataset at the top ("from Carrier mix review"), read-only. Click Open in the chat to go back to them.
When the agent decides a request is better done somewhere else, it shows a card that says "Your request goes with you" (or "Your request and what I found go with you"). Pick Continue in Automations or Continue in the agent to take the request there, where it runs at once. The chat that receives it says where it came from, for example "Handed over from Orders". Pick Add it here to keep working in the chat you're in.
Start a new chat
Click New chat at the top of the chat list, or press ⇧⌘O (Ctrl+Shift+O). A new chat is the message box with five hints under it. Click a hint to see what you can ask:
| Hint | Example tasks |
|---|---|
| New dataset | The Data Library's ways to add data. Pick one to open the New dataset window. |
| Explore | "What stands out in a dataset?", "What changed in a dataset since last week?", "Break a dataset down by its main categories" |
| Build a dashboard | "Build a dashboard from a dataset", "A one-page summary of a dataset's key numbers", "Add a chart of a dataset to one of my dashboards" |
| Bring data together | Add one dataset's columns to another, put two datasets' rows into one, or combine a file's tabs |
| Clean | "Remove duplicate rows from a dataset", "Fix the date formats in a dataset", "Tidy inconsistent spellings in a dataset" |
Pick a task and the data picker opens. Choose the dataset and its name goes into the sentence, for example "Remove duplicate rows from orders.csv". Edit it if you like, then send.
Choose the project
A chat belongs to one project (in Ask, it can also cover All my projects). The chip under the message box says which, for example "in Sales", with what the agent may do there:
| Chip | What the agent may do |
|---|---|
| Build | Read and change data in a project you can edit |
| Explore | Read and answer, and look through rows. It changes nothing. |
| Ask | Answer only |
Before your first message, click the chip to pick another project under Ask about (a search box appears when you have more than six). "A chat stays in the project it starts in." After the first message the chip shows a lock: "This chat stays in Sales. Start a new chat to ask about another project."
With no project yet, the chip reads "No project yet". Click Create one to make one, if you're allowed to.
The message box
Type in the box ("Ask about your data, or tell Mammoth what to do") and press Enter to send. Shift+Enter starts a new line. A message can be up to 20,000 characters.
| Control | What it does |
|---|---|
| + (Add data) | Adds a file, a dataset, a connection or other data to the message. See Add data to a message. |
| Project chip | The project the chat is in. See Choose the project. |
| Intelligence chip | How much thinking the agent does, for example Fast. See Intelligence. |
| ⋯ (More options) | Choices that apply to every chat. See More options. |
| Send (arrow) | Sends the message. Its tooltip says why it can't send yet, for example "Type a message to send" or "Waiting for files". |
| Stop | Shown while the agent works. Stops the current reply (its tooltip says "Stop generating"). |
Under the box: "AI can make mistakes. Verify important results."
Send while the agent works
You don't have to wait. While the agent works, Send says "Send after the current answer". Your message waits above the box, marked Queued, and goes when the answer ends. If the agent can't take it, you see "Couldn't send that while the Agent works. Your message is back in the box." If you click Stop, waiting messages go back into the box.
If you send while a file is still being added, the message waits for it: "Sends as soon as your files are in. Edit it to hold it back."
A password in your message
If you paste connection details with a password, for example a database connection string, Mammoth offers: "This has a password in it. Put these in the connection form instead, so the password stays out of the chat?"
| Option | What it does |
|---|---|
| Put in the form | Moves the details into the agent's connection form. The message keeps only where to connect, for example "Connect to the sales Postgres database on db.example.com". |
| Send as typed | Sends the message as it is, password included |
Intelligence
Click the intelligence chip to pick a level. It applies from your next message.
| Level | Hint shown |
|---|---|
| Fast | Quick answers and simple changes |
| Balanced | Everyday analysis and pipelines |
| Expert | Hard problems; slower, thinks longer, can take several minutes |
| Auto | Plans with Expert, works with Fast |
A level the Workspace Owner turned off shows "Turned off by your workspace". Owners set the default and the levels on offer. See Intelligence.
More options
⋯ opens For every chat:
| Option | What it does | Default |
|---|---|---|
| Auto-approve plans | "Run plans and changes without asking." Off, each plan waits for your OK. | Off |
| Alert when work finishes | "A desktop notification, while you're away." Mammoth asks your browser for permission the first time. | Off |
If your browser blocks notifications, you see "Your browser blocks notifications from Mammoth. Allow them in its site settings." For how plans and approval work, see Plans and runs.
Add data to a message
Click + (Add data) in the message box. The tabs are:
| Tab | What you do |
|---|---|
| Library | Pick a dataset or view from the chat's project ("Search datasets and views in Sales…"). It lists Recent ones; type 3 letters or more to search. Add puts it on the message as a chip. Browse all… opens the full picker. Shown when the project has data. |
| Upload | Drop a file, or click Choose a file. "CSV, Excel, PDF and more. ⌘U opens this from the message box." |
| Connect | Pick from Sources or Your connections ("Search sources…"). The agent starts connecting it, for example with the message "Connect HubSpot". |
| More | Paste rows ("Rows from a spreadsheet or an email"), Public data ("Census, weather, prices and more") and Sample data ("Ready datasets to try things on") |
Upload, Connect and More are shown only to people who can add data to the workspace.
| Option under More | What happens |
|---|---|
| Paste rows | Paste rows with the header row first, tab or comma separated, and click Add rows. "The rows upload as a CSV and go with your next message." |
| Public data | Sends "Find public data for me". The agent finds public datasets and shows each with its rows, columns and date range. Click Import, or Import last N days where offered (for example, Import last 30 days; the number depends on the source). A set too large to import says so: "Too large to import here (2,400,000 rows). Open it on data.example.gov to download a filtered part." |
| Sample data | Pick a theme, then a sample. It's added to the project and goes on the message as a chip. |
You can also drag a file onto the message box ("Drop a file to upload it") or press ⌘U (Ctrl+U).
Uploaded files
An uploaded file goes into the chat's project, and the agent picks it up from there: "orders.csv added to Sales". If the chat covers no single project, Mammoth asks "Which project should this file go to?" Pick one and click Upload here. Several files upload one at a time; each chip shows its progress, and × cancels one.
| Message | What it means |
|---|---|
| "orders.csv was flagged as unsafe and was not uploaded." | The file failed a safety check |
| "orders.csv is not a supported file type." | Mammoth can't read this kind of file |
| "orders.csv could not be uploaded. Try again." | The upload failed. Click Retry on the chip. |
| "orders.csv uploaded but could not be processed." | The file arrived but couldn't be read as data |
| "orders.csv uploaded, but the assistant is busy. Send a message to continue." | The file is in the project. Send a message to tell the agent. |
Pictures and documents
Pictures and documents you add in the message box are read by the agent, not imported. Say what each one is for (a picture's box suggests "Say what this is for…"), or the agent asks.
| Kind | Accepted | Notes |
|---|---|---|
| Pictures | PNG, JPEG, GIF or WebP, up to 2 MB | Paste, drop or upload. Large pictures are scaled down first. Otherwise: "This picture can't be used (PNG, JPEG, GIF or WebP, up to 2 MB)." |
| Documents | PDF, Word or PowerPoint, up to 20 MB, no password | The chip shows the length, for example "12 pages", or "212 pages · I'll read the first 50". Otherwise: "This file can't be read (PDF, Word or PowerPoint, up to 20 MB, no password)." |
- You can add up to 5 pictures and documents to one message: "5 pictures per message. The rest weren't added."
- A document's tables reach the Data Library only when you ask the agent to save them. See Read PDFs and workbooks.
- On each message the agent looks again at the 6 newest pictures and the 3 newest documents in the chat. Older ones are only named, so add a file again if you need the agent to read it again.
For all agent limits, see Limits and quotas.
What a reply shows
While the agent works, its reply shows what it is doing, for example "Thinking…", "Building your chart" or the command it is running. When it finishes:
| Part | What it is |
|---|---|
| Work summary | "Worked for 12s" (or "Worked for 1m 5s, in the background"). Click it to see each step and the changes made. A group of steps reads, for example, "Ran 3 commands". |
| Answer | The agent's reply in text |
| Cards and charts | Results, plans and changes the agent made. See Results and the data canvas and Plans and runs. |
| Sources | The web pages the answer used, three at a time, then "+2 more" |
| Questions | When the agent needs a choice, it asks with options. The one it suggests is marked Recommended. Other… lets you type your own answer. With several questions, Confirm sends your answers ("Answer every question to continue"). |
In a long conversation, older messages are summarized so the chat can keep going. A line in the chat says so: "12 earlier messages were summarized to keep this conversation going".
If a reply was cut off (for example, by a restart), click Continue to pick up where it stopped.
Act on a message
Hover a message to see its actions. On a touch screen they show under the message.
| Action | On | What it does |
|---|---|---|
| Copy | The agent's replies | Copies the reply's text ("Copied") |
| Try again | The latest reply | Asks the same question again |
| Edit and resend | Your own messages | Takes the chat back to that message and puts its text in the box to change and send. Later messages leave the chat. |
| 👍 Good answer / 👎 Bad answer | The agent's replies | Rates the answer ("Rating saved"). Your rating stays after you reload. |
When something goes wrong
| Message | What to do |
|---|---|
| "Reconnecting to the Agent…" | Wait. Mammoth reconnects on its own. After a while: "Can't reach Mammoth. Check your connection; we'll keep trying." |
| "Lost the connection to the Agent. Your message was received and the reply may still be on its way." | Click Reconnect |
| "This reply failed. Nothing was changed." | Click Try again |
| "The Agent hit an error." | Click Try again, or Reword your question |
| "I couldn't find an answer. Try rewording or contact support." | Click Reword and ask another way |
| "Could not stop the Agent. It is still working. Try Stop again." | Click Stop again |
| "This chat no longer exists, so it can't take new messages." | Click Start a new chat with this message |
| "Could not load chat history." | Click Try again |
The chat list
The list on the left of the Agent page holds your chats, newest first.
| Control | What it does |
|---|---|
| Search (magnifier beside Agent, ⌘K or Ctrl+K) | Search chats by title and content. A match in a message shows under the chat's title. |
| New chat (⇧⌘O or Ctrl+Shift+O) | Starts a new chat |
| Project | Shown when your chats span more than one project. Shows All projects, or the chats of one project. A project with none says "No chats in Sales yet" with New chat in Sales. |
| Round handle (⌘B or Ctrl+B) | Hide chats or Show chats. With the list hidden, New chat moves to the chat's header. |
Chats are grouped:
| Group | What's in it |
|---|---|
| Pinned | Chats you pinned, the newest pin first |
| Email and Slack | Chats started by email to agent@mammoth.io or in Slack, folded under one row with a count, or "2 need reply" |
| Today, Yesterday, Previous 7 days, Previous 30 days | Your chats by when you last used them. Older ones are grouped by month, for example "September". |
| Shared with you | Other people's chats shared with your project. You can read them; only the owner can continue one ("Shared with you to read. Only its owner can continue it."). |
A row shows the chat's title, its project (with All projects chosen), a mail icon for a chat started by email, a share icon for a chat you shared, and the owner's picture on a chat shared with you. A chat docked in a dataset or automation shows where it lives, for example "In orders.csv".
With no chats yet, the list says "No conversations yet." A search with no match says "No conversations match your search."
Status dots
A dot at the end of a row says the chat has news. Hover it to read what's going on.
| Dot | Status words |
|---|---|
| Amber | Waiting for your OK, Needs your answer, Waiting for you, Failed, Out of time |
| Blue | New result |
| Working | Working… |
To clear an amber dot without opening the chat, open the row's ⋯ menu and click Dismiss (for example "Plan waiting · clears the dot").
Chat menu
Click ⋯ on a row, or ⋯ (Chat options) in the open chat's header.
| Option | What it does | Who |
|---|---|---|
| Dismiss | Clears the amber dot | Shown when the chat waits on you |
| Pin / Unpin | Keeps the chat at the top of the list | Anyone who can read the chat |
| Rename | Edits the title in place, up to 80 characters ("Too long: a title can be up to 80 characters.") | The chat's owner |
| Move to project | Moves the chat to another project. While the chat works: "Wait for the chat to finish before moving it." Not in Ask. | The chat's owner |
| Share with project / Make private | See Share a chat with your project. Open chat's header only, not in Ask. | The chat's owner |
| Ask before running plans | "On: each plan waits for your OK" or "Off: plans run without asking". The same choice as Auto-approve plans. Open chat's header only, not in Ask. | Anyone |
| Agent settings | Opens the AI Agent page of Workspace Settings. Open chat's header only. | Workspace Owners |
| Delete | Deletes the chat, after you confirm | The chat's owner |
Share a chat with your project
In the open chat's header, click ⋯ > Share with project. The dialog asks "Share this chat?": "Members of Sales will be able to read this chat and see what it made." "Only you can continue it. You can make it private again from this menu." Click Share.
The chat then shows a share icon, and members of the project find it under Shared with you. To stop sharing, click ⋯ > Make private ("This chat is private again.").
A chat about all your projects can't be shared: "Only a chat about one project can be shared".
Delete a chat
Click ⋯ > Delete. The dialog asks, for example, "Delete “Q3 revenue”?": "The conversation is deleted for good. This cannot be undone." What the chat made stays: "Anything it made stays in your projects." A shared chat adds "Members of Sales will no longer see the chat." Click Delete chat.
The chat header
The open chat's header shows its title (click Rename in the menu to change it) and, when the chat made or changed something, a button such as "3 things made or edited". It opens In this chat, a list of what the chat made, edited and read, each with Open. See Results and the data canvas.
Chats in Ask mode
In Ask mode, your chats sit in the left navigation under Chats, with + for a new chat. Ask chats can't be moved or shared with a project. See Ask mode.
On a phone or tablet
The Agent page works on phones and tablets.
- On a tablet, the chat list opens over the chat. Tap beside it to close it.
- On a phone, tap the chat icon in the top bar to open your chats. The list fills the screen, with New chat at the bottom. Tap ✕ to close it.
- On a phone, the open chat's title, its "things made or edited" count and its ⋯ menu sit in the top bar.
- Copy, Edit and resend and Try again show under each message, since a touch screen has no hover.
- The chat on the Automations tab opens as a sheet at half the screen's height. Tap the handle at its top to Show more of the chat, and again to Show more of the page.
Continue a chat from the CLI
There is no button that moves a chat to the CLI, but the CLI can send messages to the same chats. The chat's ID is in the Agent page's address after ?session=. Pass it as session_id, with your message and scope, in the --input of mammoth agent chat to continue that chat from your terminal. The reply appears in the app too.
For coding agents and scripts, see Mammoth for coding agents and the Mammoth MCP server.