Use the workspace assistant

Team admins see Workspace assistant on Dashboard. It works in the selected Personal or team workspace.

The assistant can start Data Removal setup, update one validated profile or preference field, and submit a complete profile through the existing Data Removal workflow. Profile changes and submission require explicit confirmation in the current message. Identity values, authorization data, broker evidence, credentials, payment actions, and administrator jobs are not exposed.

You can ask the assistant to search, count, or show details for contacts. Contact search can use names, company, notes, tags, and whether an email address is set. Searches show up to 100 contacts at a time. Use Show more for the next page. Counts cover every match rather than only the displayed page.

The assistant can create one Contact after you explicitly confirm the proposed fields. A name is required; email, phone, company, tags, and notes are optional. The name is saved in the Contact's first, middle, and last-name fields while keeping the full display name. It checks the confirmation in your current message before writing, creates only the Contact requested, and returns a link to it.

The assistant can count records across Magpie product areas and has detailed read tools for Books, Calendar, Contacts, Design, Documents, Favicons, Forms, Idea Workspaces, Image Converter, Invoicing, Kanban, Links, Message Boards, Meetings, Newsletters, Notes, Podcast, Presentations, QR Codes, Reader, Screen Recorder, Spreadsheets, Store, Supporters, Team chat, To-dos, Video Chat, and Websites. Ask “Show unpaid invoices”, “List my websites”, “Which newsletter was sent most recently?”, or “Show incomplete To-dos due this week?”. Filters and result limits are handled by the relevant feature provider. Account-owned areas use the signed-in account; team features use the selected workspace. Private Calendar events are included only for their owner; other team members see team-visible events in Calendar counts and searches.

Mail appears in the capability list, but the assistant cannot count or read Mail records yet. A request for a Mail count reports that it is unavailable.

The assistant can also create a new draft Form from a built-in blank, contact, feedback, registration, request, or RSVP template. It asks you to confirm the title and template before creating the Form; it does not publish it automatically.

When a search finds a record and you ask for its notes or other stored content, the assistant can read that exact record directly. It uses the result and UUID from the search, checks access again, and reports when long content is shortened; you do not need to repeat the search or confirm an ordinary read.

For Spreadsheets, the assistant can create a blank workbook, import bounded CSV or TSV text into a new workbook, apply one supported grid operation, create or update personal Filter Views, and manage direct collaborators. It resolves exact workbook, sheet, Filter View, operation, and collaborator IDs through the Spreadsheet service. Imports, grid operations, deletion, and collaborator access changes require a strict confirmation argument and explicit confirmation in the current message. Grid writes use the current workbook revision and an operation UUID so stale or retried writes do not overwrite newer work. Cell contents are not returned in assistant evidence; successful writes return a saved result and link.

For Websites, the assistant can create, update, publish, and delete one exact team Website through WebsiteService. Website searches return metadata, not page documents. Updates use the current Website version, accept one complete bounded page document, and preserve its publication state. Publishing and deletion require a strict confirmation argument and explicit confirmation in the current message. Successful writes return a saved result and link; a provider retry reuses the write receipt.

For Presentations, the assistant can create, duplicate, update the draft, publish, change private or link sharing, add one direct collaborator, and delete one exact presentation through PresentationService. Resolve the presentation UUID from a search. Draft updates and publishing use the current draft version; slide documents are bounded and sanitized by PresentationPolicy and are not returned in assistant evidence. Publishing, changing sharing, adding a collaborator, and deletion require a strict confirmation argument and explicit confirmation in the current message. Link sharing is public and requires a published snapshot. Successful writes return a saved result and link; a provider retry reuses the write receipt.

For Designs, the assistant can create and apply bounded operations to one exact Design through DesignService. It can archive or delete Designs, add or resolve comments, manage collaborators, create or restore versions, and create, update, or merge Branches. It can also publish or update Design Libraries, disable an enabled Library, and create a public prototype share for one exact prototype flow. Resolve Design, Branch, version, comment, collaborator, Library, and flow identities from server-owned reads; titles and usernames are not destructive identities. Design operations use their operation and client-session UUIDs plus the current base revision, so stale changes are rejected or reconciled and a retry reuses the existing operation receipt. Design snapshots, operation payloads, Branch snapshots, Library manifests, and share tokens are not returned in assistant evidence. Archive, deletion, version restoration, Branch merges, collaborator changes, Library changes, and public Library or prototype sharing require a strict confirmation argument and explicit confirmation in the current message. Successful writes return a saved result and link.

For QR Codes, the assistant can create one static or dynamic code with bounded website or plain-text content. It can change the destination of a dynamic code, enable or disable a saved code, and delete one exact saved code. Resolve the QR code UUID from a QR code search; names and destinations are not identities. Payload, design, logo, and destination history are not returned. Destination changes, enabling, disabling, and deletion require a strict confirmation argument and explicit confirmation in the current message. Successful writes return a saved result and link, and a provider retry reuses the existing receipt.

For Images, the assistant can upload one bounded JPG, PNG, GIF, or WebP image to the signed-in account's Image library and soft-delete one exact library image. For Favicons, it can create or discard a transient package for one exact Favicon project. For Image Converter, it can convert an exact library image, save a conversion to the library, create a bounded batch from exact library images, or discard one conversion. These operations use the existing ImageService, FaviconGeneratorService, ImageConversionService, and batch worker boundaries, including their normal quota and expiry rules.

Resolve the exact Image, Favicon project, or conversion UUID first; filenames and project names are not write identities. Every media write requires a strict confirmation argument and explicit confirmation in the current message. The assistant returns saved results and server-issued links only. It does not return image bytes, temporary or storage paths, transform JSON, package paths, or website image tokens. Image deletion is soft, and discarding a Favicon package leaves its saved project and source image unchanged.

For Screen Recorder, the assistant can start a capture, append the next bounded recording part, finalize it, rename it, update its share-page settings, save a non-destructive cut plan, and delete one exact recording. It can also create, update, or delete exact image overlays on a ready recording. Resolve recording, overlay, image, and transcript Document UUIDs from current Magpie results; titles and filenames are not identities. Sharing, public exposure, edit plans, overlay changes, and deletion require a strict confirmation argument and explicit confirmation in the current message. The existing recording services enforce plan, chunk, storage, timing, geometry, expiry, and cleanup rules. The assistant returns bounded metadata and server-issued links only: it does not return video bytes, private paths, share tokens, image filenames, credentials, or transient video signaling.

Each feature says which text a search reads. A search of Notes reads the title and the Markdown body; a search of Kanban cards reads the card title, the card description and the board name; a search of Form responses reads the form title and the respondent's display name; a search of To-dos reads the title and the notes.

Content inside a Book, a Podcast or a links page has its own search. Ask for what a Book says and the assistant searches its sections — titles, synopses and section text — and names the Book each match came from. Podcast episodes are searched by title and show notes; the audio is never read, so spoken words are not searchable. Link blocks are searched by label, subtitle and destination. Searching the Books, Podcasts or links pages themselves does not reach inside them, and the assistant says which of the two it did.

Comments can be found by their own words. Ask who mentioned something and the assistant searches Kanban card comments and Message Board comments across the workspace, naming the card or post each one sits on — you do not have to find the card or the post first. A result names the fields it searched and, for each row, the field that matched, so "nothing found" and "that is not searched here" stay different answers. Ask the assistant what it searched if a result looks narrower than you expected.

Some searches read no free text at all. Video rooms hold no text, so the assistant does not offer a word search for them rather than offering one that can never match.

Long text comes back as a preview, not as the whole record. The search itself still reads the whole stored text, so a word further into a long Note is found even though the preview stops before it. Where a full-content read exists, the result names it: a Note's body is read in full by UUID.

Record searches are saved with their exact matching total, applied filters, workspace or account scope, retrieval time, and a bounded preview. The preview may contain fewer rows than the exact total. If a later assistant response fails, the completed search remains listed in the conversation.

When a detailed feature search has more rows than its preview, the saved result keeps the applied search and can load the next 100 rows with Show more. Each page checks the current workspace access again. A saved preview that cannot be paged is still marked incomplete, so it is never presented as a complete list.

When a result includes a table, the reply shows a short View link with the number of available rows. Open that link to reveal the table. Large results keep Show more inside the revealed table and update the displayed row count as more records load.

This saved evidence also applies to the detailed readers for meetings, repositories, Reader entries, Message Board posts, Store orders and products, Supporters, To-dos, Documents, and Team chat.

An exact product-area count is also shown from its saved result. It remains visible even when the assistant's written response omits the result reference.

For Idea Workspaces, the workspace is the canvas and its items are the notes, to-dos, links, images, comments, columns and connectors inside it. Ask the assistant to find canvases by title, search their items, or count items across selected canvases. Item counts are calculated for all matching items and can be grouped by canvas or item type. A displayed item list is a bounded preview; the assistant reports separately whether the preview and requested canvas scope are complete.

Idea Workspace discoveries, item searches and counts are saved with the conversation. A follow-up such as “only the images in those canvases” uses the saved canvas selection but reads current records and checks access again. Ask to use the earlier result when you want the saved snapshot instead of a refreshed count. A saved result does not provide access to a canvas.

The assistant can also create and delete Idea Workspace canvases and create, update, and delete the notes, to-dos, links, images, comments, columns, and connectors inside them. Item content is type-specific. Deleting a canvas or an item is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

The assistant can also create the selected team links list, update its profile and appearance, publish or unpublish it, add and manage its blocks, and add, update, and delete social links. Publishing or unpublishing the list and deleting blocks or social links is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

To-do searches can also filter by priority, assignee, List, or tag. Results include the scheduled date, deadline, completion state, bounded notes, recurrence details, List and assignee names, checklist items, and tags. A date range matches either the scheduled date or the deadline. The assistant can also create, update, complete, reopen, move, snooze, set recurrence or reminders, change tags, and manage checklist items on one exact accessible To-do. It can create, update, archive, and manage collaborators on To-do Lists. Deletes, bulk changes, List archive, checklist-item deletion, and collaborator access changes require an explicit confirmation in your current message. Writes use the normal To-do services and return a saved result and link; a failed provider retry reuses that receipt instead of repeating the write.

Store reads include safe product catalog metadata plus order payment and fulfillment status. Supporter reads include program, tier, subscriber display name, and subscription status. Neither feature returns Stripe identifiers, payment credentials, addresses, access tokens, or other secret values. Forms can report response counts and submission activity without returning answer payloads or uploaded files.

A complete saved Store order selection can be followed to its immutable line-item snapshots. The assistant returns product and variant names, SKU, quantity, currency, and integer minor-unit amounts. It does not return addresses, payment identifiers, access tokens, or credentials.

The assistant can follow a saved Contact selection to Invoices linked to those contacts. Ask it to find the contacts first, then ask which draft, due, overdue, paid, or void invoices belong to them. The relationship uses the contact stored on each invoice; customer names are not used to guess a match. Invoice results can also be limited by an inclusive issue-date range.

A complete saved meeting selection can also be followed to invoices linked to those meetings. This uses the invoice’s stored booking relationship and checks the meeting owner and selected workspace again; attendee names and dates are not used to guess an invoice. The result can be filtered by draft, due, overdue, paid, or void status and by an inclusive issue-date range.

A complete saved Form selection can be followed to its Responses. Response metadata can be filtered by new, reviewed, or archived status and by an inclusive submission-date range. The assistant returns the Form, respondent display name, status, and submission time. It does not return answers, uploads, or payment data.

A complete saved Newsletter selection can be followed to its posts. Post metadata can be filtered by delivery status and by an inclusive creation-date range. The assistant returns the subject, delivery totals, status, dates, and a link to the saved post page; subscriber addresses, delivery details, and message bodies are not returned.

Notes and Documents support bounded text search. A Note or Document can be retrieved by UUID as bounded plain text. Team chat can search accessible messages or return recent messages; direct-message membership is checked for every result. Git repository reads are limited to data stored in Magpie: branches, recent commits, the file tree, README text, and an explicitly requested bounded text file. The assistant does not access GitHub, GitLab, remote credentials, or unconnected platforms.

The assistant can create and update Notes in the selected workspace. Updates use the Note revision that was read and stop when another device changed it. It can also convert one unchecked Markdown checkbox into a related To-do through the normal Notes and To-dos flow. The durable assistant receipt prevents that write from repeating after a retry or reload. Deleting a Note is destructive: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

The assistant can also update and delete Documents in the selected workspace. Updating Document content uses the revision that was read and stops when another device changed it. It can add and remove collaborators, add comments, add and decide suggestions, and create or restore versions. Use the Document collaboration read to resolve the exact collaborator, suggestion, version, and comment IDs before those writes. Deleting a Document, changing collaborators, deciding a suggestion, or restoring a version is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

For a question about themes, similarities, sentiment, or another content-based conclusion, the assistant reads bounded Note, Document, Reader, or chat content before answering. These conclusions are inferences from the content it reviewed, and the reply says when it saw only a bounded preview. Exact counts and stored relationships continue to come only from the relevant Magpie result.

Use “What can you do here?” to retrieve Magpie's current feature and capability catalog. Product areas publish their assistant operations through capability providers. Each provider owns its tool descriptions and schemas and calls the same application services used by Magpie's other interfaces. Registering a provider makes its tools available to the next OpenAI request and adds them to the live administrator reference. Duplicate provider keys and tool names are rejected. Password Vault plaintext is excluded because it is encrypted and is not available to the server.

You can also list, search, summarize, create, update, and delete Kanban boards in the selected Magpie workspace. Board creation uses the normal Inbox, To Do, In Progress, and Done columns and team visibility. The assistant can also create, rename, move, and delete columns, create, update, move, delete, and vote on cards, and add or delete card comments. Resolve a board before a board-scoped write, and use the column ID returned by the board summary and the card ID returned by a card search. Deleting a board, column, card, or comment is destructive: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself. Column deletion is limited to empty, non-Inbox columns. An unqualified reference to “my board” means a Magpie board here. The assistant asks which Magpie board only when more than one board remains possible; it does not ask which outside service you use.

Card searches can filter by board UUID or a saved single-board result, column name, or due-date range and return the card description, assignment, suggestion state, comment and vote counts, and the board link. Resolve a board with List Kanban boards before filtering cards by board, or use its saved result when one board remains. Card content remains bounded.

Reader searches can find entries by title, author, feed, or bounded plain-text content, and can filter by read state or publication date. They include the stored original URL when one exists and a link back to the entry in Magpie.

Message Board searches can find posts by title, board, or bounded body text and filter them by board slug or creation date. Results include the board, author, lock state, score, comment count, and a link to the post. Private-board results are limited to members of the selected team.

A complete saved Message Board post selection can be followed to its comments. The assistant uses each post’s stored identity, checks selected-team access again, excludes deleted comments, and returns bounded comment text with a link back to the post. Comments can be limited to an inclusive creation-date range. It does not infer comments from matching titles or authors.

The assistant can also create and update Message Boards, and create, vote on, and delete posts and comments through the normal board service. Deleting a post or comment is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

The assistant can create, update, publish or unpublish, and delete one exact account-owned Book through the normal Book service. It can also create, update, move, include or set aside, delete, and restore a version of one exact Book section. Resolve the numeric Book, section, or version ID first; titles are not identities. Publishing, deleting, and restoring a version require an explicit confirmation in the current message. Setting a section aside is reversible, and section HTML is sanitized by BookService. Successful writes return a saved result and link, and a provider retry reuses that receipt instead of repeating the write.

A complete saved Kanban card selection can be followed to its comments. The assistant follows each card’s stored identity, checks selected-team access again, excludes deleted comments, and returns bounded comment text with a link back to the board. Comments can be limited to an inclusive creation-date range; matching titles or authors are not used as a guess.

For board summaries, cards in columns named Done, Complete, Completed, Resolved, Closed, Finished, Shipped, Archive, or Archived count as resolved. Cards in every other column count as unresolved. The result names the columns used for the resolved count.

The assistant can create a new Document from text you provide. It can also use a saved contact result. For example, ask it to find iOS developers and then say “Put those people in a document.” The document uses the full saved result, including contacts beyond the first page. A contact selection can contain up to 5,000 contacts and 2 MiB of projected data. Narrow a search that exceeds either limit.

You can also ask it to put a saved count, comparison, or sum in a Document. These calculation sources are added as a short summary; raw record rows are not copied through this path.

Ask the assistant to create a meeting with a person's name, date, and time. Add a length when it should not use the 30-minute default. An email address or phone number is optional. Magpie first looks for one matching contact in the selected workspace, using an exact email or normalized phone number when supplied. It reuses that contact or creates a CRM contact, then creates the meeting. If more than one contact matches, the assistant asks for a more specific email address or phone number.

Creating the meeting does not initially send email or add Google Meet or Zoom. The assistant offers those choices after the meeting exists. If you accept one, the follow-up uses the saved meeting ID. Google Meet requires a connected Google Calendar account, Zoom requires a connected Zoom account, and sending email requires an attendee email address. Adding Google Meet by itself does not send a Google Calendar invitation. Assistant email remains subject to the simulation setting described below.

The assistant can also reschedule, confirm, cancel, archive, and delete a Meeting by its UUID. Rescheduling uses the existing conflict check and does not send meeting email. Cancelling or deleting a Meeting is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

The assistant can create, update, and delete your Meeting types and your weekly availability windows. Deleting a Meeting type or availability window, and clearing all availability, is consequential: the assistant asks for an explicit confirmation in the current message.

The event campaign action searches contacts or uses a saved contact result. It creates an RSVP calendar event, publishes an attendance and food-and-drink Form, adds a Zoom meeting when requested, and prepares a single-use invitation for each matching contact with an email address. A campaign can prepare up to 100 unique email recipients. Events must be in the future and within the next 89 days.

Include the event date, start and end time, location, whether it needs Zoom, and the food and drink choices. The assistant asks for missing required details before creating anything. Connect Zoom from Integrations before requesting an online meeting.

Invitation email delivery is simulated by default. Magpie creates the invitation records and reports how many messages it would send, but sends no email. The simulation count is recorded in the application log without contact email addresses. Keep WORKSPACE_ASSISTANT_SEND_EMAILS=false while using simulation.

The assistant can also create, update, and delete Calendar events in the selected workspace. Updating an event uses the updated_at value that was read and stops when another device changed it. It can create and revoke shareable invitation links for an event. The assistant creates invitation links but does not send invitation email. Deleting an event, creating an invitation link, or revoking an invitation is consequential: the assistant asks for an explicit confirmation in the current message and does not accept a tool argument by itself.

Conversation history and saved selections are stored by Magpie for your account and selected workspace. They remain available after a reload. Clear conversation removes the visible history and makes its saved selections unavailable. It does not delete documents, events, Forms, Zoom meetings, or invitations already created.

Each browser window keeps its own current conversation. Use Open new conversation to open another Dashboard window with an empty conversation. Both windows can run requests at the same time because each uses a different server conversation. A single conversation still runs one request at a time.

Use the new-conversation and clear-conversation icons above the chat, or type /new to start a blank conversation in the current window and /clear to clear the current conversation and its saved selections. Links returned in an assistant response open in a new browser tab.

When a result includes a table, record links appear in that table instead of a second list above it. If the written answer already states the exact total, the table or saved count is shown without repeating the total in another line. Internal result field names are not shown in the conversation. Fallback summaries use the record name for every supported product area. When a grouped count already names every group and its value in the written answer, the duplicate table is omitted and the group links remain available. The same rule applies to a saved-count sum that already states every source total.

The assistant can compare two or more exact counts already saved in the current conversation. The comparison uses those saved snapshots and shows each total, the difference from the lowest total, and the percentage above it. Ask it to read the records again first when you need current values. A percentage is omitted when the lowest total is zero.

The assistant can also add two or more saved exact counts from different product areas. The result shows the arithmetic total and each source count. This operation does not remove duplicate records from overlapping selections; use a saved-list union when same-type records need to be deduplicated.

Saved read results also work across product areas. You can ask for an earlier snapshot without refreshing it, or combine complete saved lists of the same record type. Supported list operations are union, intersection, and difference. Saved snapshots are read in pages of up to 100 rows. Combined lists are limited to 100 records and require a stable record UUID or ID.

Each request has a fixed limit of six OpenAI responses and twelve tool calls. If a request stops or reaches a limit, completed records remain listed. Use Resume after a network interruption. Magpie reuses the saved request and write receipts instead of starting the same work again.

For a request that uses more than one product area, the working reply names the current operation and the record type being read. If one operation fails, the final reply labels the completed, failed, and unfinished parts while retaining results from the parts that completed. These labels use product-area names and do not show internal tool names or arguments.

While a request runs, each operation update remains in its own assistant reply instead of being replaced by the next update. A completed read includes its exact record count when one is available, and completed documents, meetings, event campaigns, or saved calculations name the completed operation. An animated three-dot reply remains at the bottom until the assistant finishes or the request stops. Saved progress replies return with the conversation after a reload. Magpie saves and displays an operation before executing it, so its pending reply can appear before the completed result.

Compound requests also keep a task checklist with dependencies between steps. Step status comes from completed tool calls rather than generated status text. If the assistant tries to finish while an independent step is still pending, Magpie asks it to continue the saved checklist. This continuation is bounded by the normal request limits. A step whose required source failed is marked blocked.

A read that stops because its data source is temporarily unavailable is tried once more within the same request. Validation, permission, and unsupported-operation errors are returned without a retry. Write operations are never repeated through this read-recovery rule.

An active request uses a short execution lease. If a different request or Clear conversation finds that the active run has made no progress for five minutes, Magpie closes the inactive run and releases the conversation. Known result links remain in its saved response. Retrying the original request before replacing it resumes the existing run.

Magpie sends the instruction, the conversation items needed for the follow-up, the current date, account timezone, relevant contact fields, bounded Idea Workspace item text, and tool results to OpenAI. Raw image identifiers and full link addresses from Idea Workspace items are not included. A contact's notes are sent only when a tool request needs them. Magpie does not send account credentials, API keys, or billing fields. OPENAI_API_KEY stays in the server environment, OPENAI_MODEL selects the model, and API response storage remains disabled with store: false. OPENAI_REASONING_EFFORT can set a reasoning level supported by that model; leave it blank to use the model default. The active value appears in the administrator assistant reference.

Request timing is recorded in the application log for the request, run lookup, lease, OpenAI response, tool execution, and progress-save stages. Set WORKSPACE_ASSISTANT_AUDIT_OPENAI=true to also record readable multiline copies of the semantic input sent to OpenAI and the output it returns. Audit entries omit request headers, credentials, CSRF values, and encrypted reasoning content. They can contain instructions, contact details, and document text, so leave this setting off unless the application log is being reviewed and protected for that purpose.

Website administrators can open Admin → Assistant reference to review the exact current system instructions, provider settings, registered tool descriptions and argument schemas, final-answer schema, limits, and execution flow. It also shows which capability provider owns each tool. The tables read from the same registry sent to OpenAI, so they change when a provider is registered or one of its tools changes.

Workspace membership is checked when history, results, and tools are used. Reads use workspace view access, and account-owned reads use the signed-in account. Password Vault remains unavailable because its contents are client-side encrypted; the assistant cannot read or unlock it. Data Removal exposes only your workflow state and aggregate request summaries, never removed content, encrypted identity data, broker evidence, or internal queue details. Creating a document, meeting, meeting option, or event campaign requires edit access. For a result that does not match the instruction, Contact support.