Assistants

Assistants are the governed AI agents that power Threads — one per team or process (legal, finance, sales, support). Each Assistant runs on any model you choose, connects to your data, knowledge, tools, and skills, and is built entirely no-code. They can be shared across all projects in the organization or scoped to a single project, with access controlled per user and group.

Assistants are managed under Settings → Assistants in MINEO. Each Assistant is configured through a tabbed form with the following sections: General, Capabilities, Knowledge, Integrations, Skills, Permissions, and Usage. File and service integrations (Google Drive, SharePoint/OneDrive), personal accounts (Gmail, ...), and MCP servers live under Integrations.

Managing Assistants

The Assistants list shows all Assistants available in the project. From here you can:

• Search Assistants by name or description.
• Create a new Assistant with the + button, or from the Assistant selector of a new thread (New Assistant). See Creating an Assistant.
• Set a default — tick the Default checkbox to make an Assistant the pre-selected option when creating new threads. Only one Assistant per project can be the default.
• Clone an Assistant from its context menu to duplicate it as a starting point. The copy keeps the configuration, allowed models, skills, conversation starters, MCP servers and uploaded attachments, is named "<name> (1)", and is never the project default. Cloning a shared Assistant always produces a project-private copy.
• Delete an Assistant from its context menu. An Assistant that is still linked to existing threads cannot be deleted.

Creating an Assistant

The + button opens a short New Assistant dialog with the essentials: Avatar, Name, Model, and Instructions. Press OK to create the Assistant and open its full form, where the rest of the configuration lives. The same dialog is available from the Assistant selector of a new thread (New Assistant).

Quick create from the name

Type a name and click the magic wand next to the Name field (Quick create assistant from name). After you confirm, MINEO writes a short Description (unless you already wrote one), picks an emoji avatar (unless you already chose an avatar) and writes the Instructions from the name — improving them instead if you already wrote some — and adds up to four conversation starters, then creates the Assistant and opens its form so you can review and edit everything. The dialog stays locked while generating; closing it before confirming creates nothing.

info

Quick create is an AI request: it is checked against your AI cost limits and its token usage and cost are recorded like any other model call. If the quota is exhausted, the dialog shows the quota error and keeps what you typed.

General settings

Field	Description
Avatar	An image that represents the Assistant in the thread UI.
Name	The display name shown in the sidebar and thread header.
Default model	The LLM used to answer when a thread sends no per-message override.
Allowed models	The set of models a user can switch between from the chat input. The default model must be one of them. This is what makes Threads model-agnostic — pick the best model per task and switch providers without rewriting the Assistant.
Description	A short text describing what the Assistant does.
Instructions	A system prompt written in Markdown that shapes the Assistant's behavior, persona, and constraints.

When you open the Default model or Allowed models selector, a detail panel describes the highlighted model: its provider, Capabilities (reasoning, attachments, image analysis, web search, code execution), and price per 1M tokens. A Usage consumption gauge positions the model on a Low–Moderate–High scale using its absolute cost index across the whole MINEO catalog (0–100, logarithmic, token prices weighted 3:1 output over input), so the same model reads the same wherever you see it. Hover the gauge to see the exact input and output price.

Next to the Instructions field, use Optimize (magic wand) to have MINEO rewrite your draft into a clearer, better-structured system prompt; Undo restores the previous text.

tip

The Instructions define the Assistant's task and take priority over each user's personal custom instructions, so an Assistant stays on-task regardless of who chats with it.

Conversation starters

Starters are suggested prompts shown to the user when they open a new thread. They act as quick-start buttons that send a predefined prompt automatically.

Each starter has:

• Icon — a visual icon shown on the button.
• Name — a short label for the button.
• Description — an optional subtitle.
• Prompt — the text sent to the Assistant when the user clicks the starter.

Up to 4 starters can be added per Assistant.

Click the magic wand next to Add to generate them with AI from the Assistant's name, description and instructions. You can add an optional hint about what they should cover; the generated set replaces the current starters. This is an AI request, checked against your AI cost limits like any other model call.

Input mode

Input mode controls how a new thread with this Assistant starts:

• Text (default) — the user begins with a free-text message box.
• Form — the user fills in a form before the first message. The submitted values are sent as structured text to the model at the start of the conversation, giving the Assistant a structured starting point instead of making it guess what it needs.

When Form is selected, a field editor appears below the toggle. For example, a weather Assistant can ask for City (text), Unit (select: Celsius/Fahrenheit), and Forecast days (number). The model receives the values as:

**City**: Madrid
**Unit**: Celsius
**Forecast days**: 5

Each field has the following settings:

Setting	Description
Name	The label shown next to the field.
Type	One of: Text, Multi-line text, Select, Multiple select, Number, Email, Checkbox, Date, Date range, File.
Width	How much of the row the field occupies: Full width, Two thirds, Half, or Third — so related fields can sit side by side.
Required	Toggle — prevents submission if the field is empty.
Placeholder	Hint text shown inside text-like fields.
Default	An optional default value pre-filled in the form.
Options	A list of options shown when Type is Select or Multiple select.

Some types add their own constraints: Number has Min / Max / Step; Date and Date range have an earliest and latest allowed date; File has allowed extensions, a maximum size, and a maximum number of files (up to 5). Uploaded files are attached to the first message of the conversation.

Fields are ordered and can be re-arranged with the up/down arrows.

Use the Preview button in the field editor to see the form as the user will, and check the structured text the Assistant will receive from a sample submission.

Generate with AI

Use the Generate with AI (magic wand) button in the field editor to build or edit the form from a plain-language description — for example "A support request form: full name, email, priority (low/medium/high) and a message". The AI returns a set of fields (name, type, required, placeholder, and options for choice fields) that replace the current ones, ready to fine-tune by hand.

info

Generating a form is an AI request: it is checked against your AI cost limits and its token usage and cost are recorded like any other model call.

Capabilities

Capabilities extend what an Assistant can do beyond text generation. They are configured in the Capabilities tab of the Assistant form.

• Web search — allow the Assistant to search the web for up-to-date information.
• Code execution — allow the Assistant to write and execute Python code to solve tasks or process data. Recommended for heavy spreadsheet analysis: OpenAI models read attached spreadsheets natively as a summarized view, but detailed aggregations, joins, and charts work best through the code interpreter.
• Image generation — allow the Assistant to generate images with a dedicated image model.

Image generation

When Image generation is enabled, the Assistant can generate images during a conversation. Images are always produced by a dedicated image model — never by the Assistant's chat model — so the capability works with any Assistant model (OpenAI, Anthropic, Google…).

The Generation model selector lists the image models of the catalog — OpenAI's GPT Image family (gpt-image-2, gpt-image-1.5, gpt-image-1, gpt-image-1-mini) and Google's Gemini image models (Nano Banana 2, Nano Banana Pro, Nano Banana, Gemini 3.1 Flash Lite Image) — Each model's card shows the price of one image at low / medium / high quality, a usage gauge on the same absolute scale as the other pickers (computed among image models from the price of one medium-quality image), and what the model supports (image editing from an input image, web search grounding, maximum output resolution). Adding or removing options is a matter of catalog configuration — no code change is needed. Image models can't hold a conversation, so they never appear as the Assistant's model or in its allowed models — only here. If no model is stored, the first image model offered by the plan is used; if the plan offers none, the form warns you and the capability has no effect.

Image quality (low / medium / high, default medium) sets the per-image price on models that support it: OpenAI GPT Image models render at that quality (roughly ×4 from low to medium and ×15 from low to high), and Nano Banana Pro / Nano Banana 2 render at 2K on high. Other Gemini image models always render at 1K.

The generated image is shown in the reply and handed back to the Assistant model, so it can describe it or build on it in the following turns.

Images can also be edited or combined: ask the Assistant to modify an image you attached in the current message ("put a hat on this photo") or the image it has just generated ("now make it black and white", "merge these two"). The Assistant passes those images — up to four per request: the image attachments of the current turn and/or the last generated image — to the image model together with your instructions. Older attachments and external URLs are never used. Since the image model does the rendering, the Image generation entry of the tools menu in the chat can be forced with any Assistant model.

info

Each image is a separate AI call (surface thread_image_generation): the image model reports the tokens it actually consumed, so the recorded cost is the real cost of the image. It is attached to the message that requested it and charged against your AI cost limits like any other model call.

Web search

When Web search is enabled, the Assistant can retrieve up-to-date information from the web during a conversation.

Search effort controls how much the Assistant searches and reads per response — higher profiles give better-sourced answers but cost proportionally more:

Profile	Behavior
Light (default)	Quick lookups: few searches per response, small result snippets.
Standard	Balanced: a few more searches per response.
Deep	Thorough research: more searches per response with large result snippets.

info

Each provider honors the knobs it supports. The searches-per-response cap applies to Anthropic models; the result snippet size is passed to OpenAI models that support it. Google Gemini does not expose either.

Domain restrictions let you control which websites the Assistant is allowed to search:

Setting	Description
Allowed domains	When set, searches are restricted to these domains only. Leave empty to search the entire web.
Blocked domains	Domains excluded from all search results, regardless of the query. Leave empty for no exclusions.

Enter bare domain names — for example wikipedia.org or docs.python.org. Do not include a scheme (https://) or a path (/some/page).

Provider support and behavior

Domain filtering is delegated entirely to the model provider — MINEO passes the lists in the API request and the provider applies them before returning results. This means:

• Filtering is transparent to users. The conversation flows normally; results from restricted domains are simply absent. No error or notice is shown in the thread.
• Support varies by provider. Anthropic, OpenAI Responses, Groq, and xAI honor domain lists natively. Google Gemini does not support domain lists — when a domain policy is configured (assistant or organization level), web search is disabled entirely on Gemini models rather than silently searching the full web.
• Anthropic only supports one list at a time. If both Allowed domains and Blocked domains are configured simultaneously, the Anthropic API will return an error at runtime. The assistant form warns you when both are set. Use one or the other with Anthropic-backed models.

tip

Organization administrators can configure organization-wide domain defaults in the AI tab of the organization settings. Per-assistant settings are overridden when an organization-level list is configured — the assistant form shows the active list as read-only in that case, with a link to the organization settings.

Charts

Every Assistant can render interactive charts directly inside the conversation. When the data lends itself to a visual representation, the Assistant automatically generates a chart in the thread.

Supported chart types: line, column, bar, area, pie, scatter, funnel, heatmap, gauge, histogram, and waterfall.

tip

Charts work especially well when combined with a Data source. Ask the Assistant to query your data and visualize the results — it will pick the most appropriate chart type automatically.

Emails

When Generate emails is enabled, the Assistant can compose an email and render it as an editable draft card in the conversation. You review and edit the recipients, subject, and body before anything goes out — the Assistant never sends on its own.

From the card you can:

• Send via Gmail — send through your connected Gmail account. Requires connecting Gmail under Personal accounts; if it isn't connected, the card offers a Connect Gmail action.
• Open in email app — open the draft in your device's default mail client with the subject and body pre-filled. No connection needed.
• Copy — copy the draft to the clipboard.

info

Composing drafts and sending through Gmail are independent. The card's Gmail send uses your own personal connection, so an Assistant can draft emails even with no Gmail tools enabled on it.

Knowledge

The Knowledge tab groups everything that gives an Assistant access to external data and files.

Data source

Connect a datasource to allow the Assistant to query and analyze it using natural language. This is the primary way to give the Assistant structured data to work with.

Attachments

Upload files that the Assistant can reference in the conversation. Supported file types include documents, spreadsheets, images, and other common formats. Each Assistant has a configurable size and count limit for attachments.

File system

Setting	Description
Working directory	Select a folder in the project filesystem so the Assistant can explore, read, and analyze files.
File write access	When a working directory is set, enable this option to allow the Assistant to create, edit, move, copy, and delete files inside it. Disabled by default.

info

Connect external file services (Google Drive, SharePoint/OneDrive), personal accounts (Gmail, ...), and remote MCP servers under Integrations.

Skills

Skills are curated prompt fragments that attach specialized instructions to an Assistant, nudging its behavior in a specific direction. Each skill has a loading mode that controls when its instructions reach the model — either always in the system prompt, or loaded on demand when the model decides it needs them.

Skills are grouped into categories to make them easier to browse: Writing, Data, Coding, Creative, Communication, Productivity, Product & Strategy, Sales & Marketing, Finance & Legal, and People & Ops.

Enabling skills

Open the Skills tab in the Assistant form. From there you can:

1. Filter by category — pick one or more categories from the selector at the top.
2. Search by name or description.
3. Click a card to enable or disable the skill. A check mark appears in the corner of every enabled skill.
4. View instructions to see the exact prompt fragment a skill will inject before enabling it.

You can enable multiple skills on the same Assistant — their instructions are combined additively.

Loading modes

Every skill is tagged with one of two loading modes:

Mode	Behavior
Always included	The skill's instructions are always part of the Assistant's system prompt. Use for behavior that should apply to every message.
On-demand	The skill is exposed to the model as a capability it can load when relevant. The instructions stay out of the prompt until the model pulls them in, so they cost nothing on messages that don't need them.

The detail view of each skill shows its mode as a badge — Always included (pin) or On-demand (download). For organization skills, you choose the mode with the Loading mode selector when creating or editing the skill.

Token budget

Each skill shows the approximate number of tokens it adds to the Assistant's prompt. The Skills tab keeps a running total of the tokens used by your selection.

warning

Always included skills consume the model context on every message, shrinking the space available for the conversation and increasing cost per message. The form shows a warning when the total exceeds a safe budget. On-demand skills only spend tokens on the messages where the model actually loads them.

Attribution

Skills imported from community repositories show their author, a source link, and the license of the original instructions. Built-in skills are attributed to MINEO and carry no source link or license.

Example skills

Skill	Effect
Creative Writer	Enhances writing with vivid imagery, literary devices, and audience‑aware tone.
Data Analyst	Adds structured methodology for querying, interpreting, and visualizing data.
Poet	Forces every response into rhyming verse.

tip

Skills are a lightweight way to experiment with different Assistant personalities without editing the main Instructions field. Combine a writing skill with a data skill to get an Assistant that explains analysis in compelling prose.

Organization-level skills

Organization administrators can create, edit, clone, and delete custom skills shared by every project in their organization. These skills are marked with an Organization badge in the skill selector. Global skills shipped with MINEO remain available to every organization and cannot be edited through the organization interface.

To manage organization skills, open the organization settings and go to AI → Skills. From there you can:

1. Add a skill — click the "+" button and fill in the name, label, description, instructions, and Loading mode (Always included or On-demand).
2. Edit — open the actions menu on an organization skill and select Edit.
3. Clone — select Clone to use an existing organization or global skill as a starting point.
4. Delete — select Delete and confirm. Global skills cannot be deleted.

Organization members and viewers can browse organization skills. Users who can edit an Assistant can select them in its Skills tab, but only organization administrators can manage the shared skill catalog. Use the scope filter (All skills / Global / Organization) to narrow down the list.

Permissions

The Permissions tab lets you control which users and groups can access the Assistant. You can grant view or manage permissions at the user and group level. This tab is only visible to users with project management rights.

Organization-level Assistants

Organization administrators can share an Assistant with the whole organization, making it available in every project.

• Share: In the project's Assistants list, open the ... menu of the Assistant and select Visibility → Organization. The Assistant then appears in every project's Assistants list and thread picker, and in the organization's Workspace → Assistants tab.
• Stop sharing: Select Visibility → Project from the same menu, or open the ... menu on the Assistant in the organization's Workspace → Assistants tab and select Project. The Assistant returns to its origin project and disappears from the other projects.
• Badge: Shared Assistants display an Organization badge in lists and in the thread picker.

The Assistant always keeps its origin project. Users in other projects can chat with it, but cannot edit, clone, delete, or set it as default — those actions stay in the origin project. Cloning it from the origin project produces a project-private copy.

info

While an Assistant is shared, every organization member and viewer can see it, regardless of the users and groups configured in its Permissions tab. The tab shows a notice while the Assistant is shared. A user invited to a project without an organization role does not see shared Assistants.

What users of other projects can reach

Sharing does not bypass file permissions. When someone from another project chats with the Assistant:

• File system (working directory): each file is checked against that user's permissions in the origin project. If a file isn't shared with them, the Assistant answers "Access denied to '<path>': this file belongs to the assistant's project but isn't shared with you. Ask a project admin to make it accessible (enable public access or grant you access)." Make the working directory public or grant access before sharing if you want it usable across the organization.
• Data sources connected in the Knowledge tab are queried with the Assistant's connection, not the user's — see Organization-level data sources.
• Attachments are part of the Assistant and are readable by everyone who can chat with it.

Attachments are shared with the Assistant

Attachments have no permissions of their own: sharing an Assistant makes the content of its attachments reachable by every organization member and viewer through the conversation. Don't share an Assistant whose attachments are confidential.

When sharing stops

The Assistant disappears from the other projects' lists and selectors, so no new threads can be started with it there. Threads that already exist in those projects are not removed and keep working with the Assistant. The organization's Workspace → Assistants tab asks for confirmation before stopping sharing.

Usage

The Usage tab shows access analytics for the Assistant: how many times it has been used, by how many distinct users, and a chart of activity over time. Filter by project, user, and period (Today, Yesterday, Last 7 / 30 days, This Month, Last Month).

This tab is only visible to organization administrators. For token spend and cost caps — as opposed to access counts — see AI Cost Limits.

See also

• Threads — start conversations with the Assistants you configure here.
• Integrations — connect Google Drive, SharePoint/OneDrive, Gmail, and other services.
• Custom instructions — personal, per-user context layered on top of an Assistant's instructions.
• Data sources — connect databases for the Assistant to query in natural language.
• AI Cost Limits — cap and monitor what Assistants spend.

🗃️ Integrations

3 items