Create an assistant

An assistant is an AI agent you set up in Nebulas. It has instructions, tools (knowledge search, web search, datatables, CRM, …), files to use as reference, and a model. Once it’s published, people can chat with it in Nebulas, and your website can reach it through nebulas-kit.

Three ways to start

In the Nebulas sidebar, click New Assistant. The main button opens the manual form. The arrow next to it opens a menu:

AI-Assisted Creation

Describe the assistant in a sentence or two. The builder agent fills in the configuration for you, and you keep refining it by chatting.

Manual Creation

Fill in the form yourself: name, instructions, tools, files and model.

Import Existing

Connect an agent that already runs somewhere else, using its A2A URL and authentication.

Which one should I pick?

If you’re not sure how to write the instructions or which tools you need, start with AI-Assisted Creation. You get the same form in the end and can change every field by hand afterwards.

AI-assisted creation

1. Click the arrow next to New Assistant and choose AI-Assisted Creation.

2. Describe what you want in the prompt box. Be specific about who uses it, what it should answer, and where the answers come from. For example:

   A customer support assistant for our website. It answers questions about billing and subscriptions using our help-center documents, stays polite and brief, and points people to the contact form when it can’t help.

3. Pick the model tier for the builder (the model that writes the configuration, not the one your assistant will use):

   • Fast (default): the quickest model that does the job well.
   • Thorough: a stronger model. It’s slower and costs more. Use it for complex assistants with several tools.

4. Send. The builder creates the assistant and answers with what it set up. When it’s done, you land on the assistant’s edit page with the fields filled in and the Agent chat still open on the right.

5. Keep refining by chatting. Ask for changes in plain language, such as “Also add web search limited to our docs site” or “Make the tone more formal”. The form is locked while the builder is working and reloads with its changes when it’s done.

6. Test it in the Test tab next to Agent. Send a few real questions and check the answers.

7. Review the form (see the fields below), then click Publish.

Note

You can reopen the builder chat on any assistant later. On its edit page, click the panel icon on the right (Get help).

Manual creation

1. Click New Assistant (or choose Manual Creation from its menu).

2. Fill in the basic info: icon, name, description and category.

3. Add tools, upload files, and write the instructions.

4. Optionally choose a model and temperature, and open Advanced.

5. Click Save draft to keep it unpublished, or Publish to make it available.

The assistant form

Both methods end up on the same form. Fields marked required must be filled in before you can save or publish.

Basic info

Field	Required	What it’s for
Icon	yes	Shown next to the assistant in lists and in the chat.
Assistant name	yes	How people find and @mention it.
Description	yes	What it does. The chat uses it to decide when to send a question to this assistant, so describe its topics clearly.
Category	yes	Groups assistants in the browser. You can pick one or click Create new category.
Declare as team resource	no	Makes the assistant available to a whole team. Only team owners and editors can do this.
Confidential	no	Only the members you select can see it.

Tools

Tools give the assistant abilities beyond the model. Click Add tools, choose one or more, and configure each (up to 10 tools). The list depends on what your organization has enabled. Common ones:

Tool	Use it to…
Knowledge base search	Answer from documents in your Nebulas collections. Set how many passages to retrieve.
Web search / Tavily web search	Search the web. You can limit results to certain domains, a time range or a country.
Web scraping	Read a web page the user or the assistant points to.
Datatable	Read and write rows in Nebulas datatables, for example to look up or record leads.
Script	Run one of your Nebulas tool scripts.
Report Template	Fill in a report template.
Image Generation	Create images, optionally following a brand’s design system.
Brand catalog	Use your brand assets.
Integrations	HubSpot, Slack, SharePoint, OneDrive, Outlook, Gmail, Google Drive, Notion, GitHub, Supabase, LinkedIn, YouTube, webhooks, and more.

Some tools (web search, for example) accept an optional alias. You can use it in the instructions to say when the tool should be used.

Files

Upload up to 20 files that the assistant can use as reference. They’re stored in a collection that belongs to the assistant.

Instructions (required)

The system prompt: who the assistant is, what it should and shouldn’t do, and the tone to use. Write it yourself or click Rewrite by Nebulas.ai to have your draft rewritten into a clearer prompt.

A good set of instructions covers:

Role:       You are the support assistant for Acme's website visitors.
Scope:      Answer questions about billing, plans and invoices only.
Sources:    Use the knowledge base search first. Never invent prices.
Tone:       Friendly, short answers (3–5 sentences), in the visitor's language.
Fallback:   If you can't answer, suggest the contact form at /contact.

Model configuration

Field	Description
Model	The model that answers. Leave it on Use workspace default unless you need a specific one.
Temperature	How random the answers are. Low (0.1) gives consistent answers. High (1.0) gives more creative, varied ones. Keep it low for support and factual questions.

Advanced

Field	Description
Recursion limit	Maximum number of steps (tool calls, reasoning) the assistant can take before it stops. Leave it empty to use the default.
Prevent query optimization	Turns off automatic rewriting of the question before searching the knowledge base.

Draft vs published

• Save draft saves your changes without making them live. If the assistant is already published, people keep using the published version until you publish again.
• Publish makes the current configuration live. The assistant’s page shows its Release status, with tabs for Configuration, Logs and Access (who can use it).

Import an existing agent

If your agent already runs somewhere else and speaks A2A, choose Import Existing. Enter its Assistant URL, pick or configure its authentication (username/password, token or OAuth), and click Preview to check its name, description and skills before you Import it. Under Advanced Settings, Force Orchestrator sends all requests through Nebulas’s orchestrator, even when the agent could be called directly.

Use the assistant on your website

The nebulas-kit chat talks to Nebulas, which then picks which assistant answers. It chooses among the assistants of the workspace that the conversation belongs to:

1. Add the assistant to a workspace. Open the workspace. In its Assistants panel, click + (Manage Assistants), check your assistant in Select Assistants and confirm. Only workspace owners and editors can do this. The assistant’s Access tab lists the workspaces connected to it.

2. Point your site at that workspace. Set NEBULAS_PROJECT_ID to the workspace’s id.

3. Make the chat create conversations in it. Pass nebulasBaseUrl to the chat so each conversation is recorded in that workspace. See Record conversations in Nebulas.

Without a conversation, the default assistant answers

Nebulas finds the workspace from the conversation. If the chat doesn’t create one (no nebulasBaseUrl, or nebulasProjectId is null in /api/a2a/config), your workspace’s assistants aren’t used and the default Nebulas assistant answers instead.

If a workspace has several assistants, Nebulas picks one based on each assistant’s description. That’s why a clear description matters.