Tools

Tools give the assistant abilities beyond conversation — finding answers in your documents, creating and converting files, reading links, remembering context, and bringing in a colleague when needed.

Most of the time you don't need to think about tools at all: describe what you want in plain language and the assistant picks the right one automatically. This page explains what each tool does, what you can ask for, and — for administrators and integrators — the technical detail and parameters behind each one. There are three kinds:

• Platform Tools — built-in abilities available out of the box.
• Frontend Tools — custom actions added by the website that embeds the chat.
• Connector Tools — abilities that come from the apps you connect.

Platform Tools

Bash

bash

Runs code in a secure sandbox so the assistant can do real work with files — generate documents, analyze data, and convert between formats. For example, it can build a Word document or PowerPoint from your outline, turn a CSV into a formatted Excel report or charted PDF, or convert a file from one format to another.

What you can ask:

• "Create a PowerPoint from these bullet points."
• "Turn this CSV into a formatted Excel report."
• "Convert this Word document to PDF."
• "Analyze this spreadsheet and summarize the trends."

The command runs inside an ephemeral Linux sandbox scoped to the chat. The filesystem persists across calls within the same chat and is destroyed when the sandbox expires. Commands run as root, and the entire filesystem is writable except the read-only Skills mounts under /mnt/skills/ and the uploaded files under /mnt/uploaded-files/. See Sandboxes for the full list of available runtimes and file-conversion tools.

Parameters:

• command - The shell command to run
• description - Short description of what this command does (5–10 words, active voice)
• timeout - Optional timeout in milliseconds. Defaults to 120000, max 600000

Chat URL

chat_url

Generates a link to open a specific chat in the app.

Parameters:

• chat_id - The UUID of the chat to generate a URL for

Memory

memory

Lets the assistant remember useful facts about you across conversations — your role, preferences, or recurring context — so you don't have to repeat yourself each time.

Memory is a per-user, per-workspace file store under /memories. The assistant views it at the start of a conversation and creates, updates, or deletes files as it learns. Paths must start with /memories, and each scope is subject to a maximum file count.

Parameters:

• command - One of: view, create, str_replace, insert, delete, rename
• path - Absolute path under /memories
• file_text - Content for the create command
• view_range - Optional [start, end] line range for the view command (1-indexed; -1 = end of file)
• old_str - Text to find for str_replace
• new_str - Replacement text for str_replace
• insert_line - Line number after which to insert for the insert command
• insert_text - Text to insert
• old_path - Source path for rename
• new_path - Destination path for rename

Present Files

present_files

Delivers the files the assistant creates back into the chat, so you can preview and download them — the finished Word document, PDF, spreadsheet, image, or any other generated file.

It downloads each file from the sandbox and attaches it to the current assistant message. Files must be within the configured attachment size limit.

Parameters:

• filepaths - Absolute file paths inside the sandbox

Semantic Search

semantic_search

Finds answers in your knowledge base and cites the source documents, so the assistant gives grounded, accurate answers from your own content instead of guessing.

What you can ask:

• "What's our refund policy?"
• "Summarize the employee onboarding guide."
• "Which document covers data retention, and what does it say?"

Under the hood it embeds the query and matches it against the workspace's document chunks by semantic similarity, returning the closest matches with citation metadata (document title, page number, and cursor) that the assistant includes in its response when citations are enabled.

Parameters:

• query - The search query. Use natural language that describes the information you're looking for; be specific and include relevant keywords
• max_results - Maximum number of document chunks to return (1–20). Defaults to 10

Send Contact Form

send_contact_form

When the assistant can't fully resolve your request, it can forward it by email to the responsible person, who will follow up with you. The assistant first asks for your contact details and your go-ahead before sending.

It emails the workspace's contact-form recipient with a summary of the issue, and can be used only once per conversation.

Parameters:

• reason - Brief explanation of the user's issue and why it requires human attention
• user_email - The user's email address for follow-up (if provided)
• user_phone - The user's phone number for follow-up (if provided)

Sleep

sleep

Lets the assistant pause and come back later in the same conversation — useful for reminders and follow-ups such as "check this again in an hour" or "remind me tomorrow morning". It uses Time to work out when to resume.

While sleeping, the chat is released from the worker and a background job wakes it at the scheduled time. For recurring follow-ups the assistant calls the tool again after each wake-up, passing iteration and total_iterations to track loop progress.

Parameters:

• seconds - How many seconds to sleep. Must be between 1 and 2,592,000 (30 days)
• iteration - Current iteration number for a recurring loop. Defaults to 1
• total_iterations - Expected total iterations for a recurring loop. Defaults to 1

Time

time

Gives the assistant the current date and time in your timezone, so it can answer time-sensitive questions and schedule reminders accurately. Often used together with Sleep.

It returns the current time in the user's timezone along with the UTC equivalent, the timezone name, the UTC offset, and the day of the week. It takes no parameters.

Tool Search

tool_search

Finds and switches on the right connector tools for a task when they're needed — for example pulling up issues from Jira or files from Google Drive — so the assistant always has the relevant abilities at hand without cluttering every conversation.

This applies only to tools marked as deferred; non-deferred connector tools are already loaded and available without a search. To keep the toolset manageable on connectors with many tools, deferred tools stay out of the conversation until this tool searches them by meaning and loads the matches so the assistant can call them.

Parameters:

• namespace - The namespace to search for tools (for example, an MCP connector)
• query - Natural language description of the capability you need
• max_results - Maximum number of tools to return (1–20). Defaults to 10

View

view

Lets the assistant open and read files and images — a document you've shared, a file it has just created, or a skill's instructions — so it can work with the contents.

It reads uploaded files, Skills mounts (paths under /mnt/skills/), and files in the sandbox. Directories are listed up to two levels deep, text files are returned with numbered lines, documents (.pdf, .docx, .xlsx, .pptx) are extracted to markdown, and images are described by a vision model.

Parameters:

• path - Absolute path, e.g. /mnt/uploaded-files/report.pdf, /mnt/skills/web-design-guidelines/SKILL.md, or /tmp/foo.txt
• description - A brief rationale for why you are viewing this file
• view_range - Optional [start, end] line range (1-indexed). Use -1 for end of file

Web Fetch

web_fetch

Reads a web page or online file and brings its content into the conversation. Share a link and the assistant can summarize it or work with the document.

What you can ask:

• "Summarize this article: https://example.com/post"
• "Pull the key figures out of this PDF: https://example.com/report.pdf"

It downloads the file and extracts its text as markdown, supporting a wide range of formats including PDFs, Office documents, HTML, plain text, and structured data files. Downloads are capped at the largest size a model can accept inline so the result is never rejected downstream.

Parameters:

• url - The URL to download the file from
• prompt - Optional instruction steering what to extract from media handled by a sub-model (e.g. images, video). Ignored for text-based formats
• timeout - Download timeout in seconds (max 120). Defaults to 60

Zendesk Escalate

zendesk_escalate

Hands the conversation over to a live human agent. Just ask to speak to a person — in any language — and the assistant connects you. It also offers to escalate when it senses frustration or a problem it can't solve.

It works in any language the assistant understands, summarizing the conversation so the agent picks it up with full context. A conversation can be escalated only once.

Parameters:

• reason - Brief explanation of why escalation is needed

Frontend Tools

Frontend tools are custom actions added by the website that embeds the chat widget — for example "add this item to the cart" or "open the booking form". They run in your browser rather than on the platform, and the embedding site defines each tool's name, description, and parameters and decides which ones are available. See Embeddable Chat → Frontend Tools for how they're set up.

Connector Tools

Connector tools come from the apps you connect, such as Jira, GitHub, Google Drive, or Slack. Each connection adds its own set of actions — reading and creating issues, searching files, sending messages, and more — which the assistant discovers and loads on demand (see Tool Search). Any MCP server can be connected; see Connectors for how they work and how to set them up.

Connectors with dedicated documentation:

• Atlassian
• Dropbox
• Exa
• GitHub
• Google Calendar
• Google Drive
• Google Mail
• Harvest
• Harvest Forecast
• Mail (IMAP/SMTP)
• Microsoft Outlook Calendar
• Microsoft Outlook Mail
• Microsoft OneDrive
• Microsoft SharePoint
• Microsoft Teams
• Redmine
• Sentry
• Slack
• Zendesk Conversations