Create your first agent
The AI-assisted builder at /agents/create-chat is the normal way to create an agent:
describe what you want in a chat panel while a live preview on the right fills in as
you talk, then create the agent when it looks right. You can also skip the chat
entirely and type straight into the preview form. Either way, nothing is saved to
your workspace until you explicitly create the agent.
Sign-in and a workspace are required
Both /agents/new (which just redirects to /agents/create-chat) and
/agents/create-chat require a signed-in session — an unauthenticated visitor
is redirected to sign in first. Creation also needs an active workspace
(tenant); see core concepts if you haven't
created one yet.
Build with the assistant
Start from a prompt, a URL, or files
On the empty builder screen, either pick a starter prompt ("Customer support agent for my business", "Lead collection bot that captures visitor info", "FAQ assistant that answers from my documents", "Feedback collection bot for visitor reviews"), type your own description, paste a website URL, or attach files from the message composer.
Pasting a URL fetches the page server-side and adds its title, description, and
extracted text to the assistant's context automatically, so you don't need to
summarize the site yourself. Attaching files accepts up to 5 files, 5 MB each, of
type .pdf .txt .doc .docx .md .json .csv .png .jpg .jpeg .gif .webp .xlsx .xls —
each is uploaded to object storage immediately, and its key is passed to the
assistant.
Answer the follow-up question
The assistant asks one or two clarifying questions based on what you gave it — what the agent should focus on for a URL, what kind of questions the files should answer, or what tone it should use for a plain description.
Review the proposed draft
The assistant replies with a suggested name, instructions, and greeting, and streams an update that fills in the live preview panel on the right in real time — you don't have to copy anything over by hand. It also proposes tools and, when your uploaded files are large or numerous, switches the retrieval strategy to RAG instead of loading files directly into context.
Edit anything directly
Every field in the preview panel is editable on its own — name, instructions, greeting, tool toggles, source URLs, retrieval strategy, mode, response limits, quick suggestions, and the anonymous-access switch. You can keep chatting to have the assistant refine things, adjust fields by hand, or mix both.
Create the agent
Two ways to trigger creation:
- Say so explicitly in the chat — phrases like "create it", "make the agent", or "go ahead and create" work; the assistant is instructed not to create on vague agreement like "sounds good" or "looks good" alone, so it will ask you to confirm first.
- Click Create Agent in the preview panel once the required fields are filled: a name (2+ characters), instructions (10+ characters), and a greeting message.
Either path inserts the agent into your workspace and, if you attached files, kicks off background processing to index them for retrieval.
Share or configure
On success you land on a confirmation screen with two shortcuts: Share Link
(/agents/:id?tab=share) and Configure (/agents/:id?tab=setup). You can also
just start chatting with the new agent directly from its page.
What the live preview controls
| Field | What it does |
|---|---|
| Name | 2–120 characters. |
| Instructions | System prompt / behavior — 10+ characters required. |
| Greeting Message | First message visitors see when they open a chat. |
| Tools | Built-in tools the agent can call: Web Fetch, File Search, Bash (shell commands over a sandboxed view of the agent's uploaded files). |
| Source URLs | Up to 5 URLs whose fetched content is kept as context; add or remove them here. |
| File Retrieval Strategy | Direct (load whole files into context), RAG (vector search on demand), or Bash (query structured files — CSV/JSON/YAML — via shell commands). See knowledge base & RAG. |
| Agent Mode | Info Provider (answers questions) or Info Collector (walks a structured field list). |
| Max responses per session / per agent | Optional caps on AI replies; blank means unlimited. |
| Quick suggestions | Off, Smart, or Always, plus how many suggestion chips (1–5) to show. |
| Allow anonymous chat | Whether visitors can chat without signing in. |
Anonymous chat is on by default
The builder pre-checks Allow anonymous chat, so a freshly created agent is reachable by anyone with its link, no sign-in required. Turn it off in the preview (or later in the agent's Configure tab) if the agent should stay private, and see public links and access gates for password- and invite-gated alternatives.
Booking managers
If you describe an owner-facing booking assistant for rooms, cabins, villas, or similar rentals, the assistant can draft a booking configuration alongside the rest of the form. That draft starts with no resources mapped — after creating the agent you still need to open its Actions tab, add each room's calendar as a bookable resource, and turn on booking before it can actually manage reservations.
Model used to build
The builder chat runs on your workspace's configured LLM provider when one is set for
the agent_creator task (see
bring your own LLM), and falls back to the
platform's OPENAI_API_KEY otherwise. If neither is available, the builder returns
an error asking you to add a provider in Settings → LLM Providers.