Quick Start
Build Your First AI Agent
A ten-minute walkthrough — create an agent, point it at a knowledge source, train it, preview it in the playground, and copy the embed snippet onto your site.
Before you start
You’ll need a Chatmount account and at least one knowledge source — a PDF, a website you control, or a few paragraphs of text. Anything the agent should be able to answer questions about.
- A Chatmount workspace (sign up at /sign-in if you don't have one yet)
- A document, URL, or block of text the agent should learn from
- About 10 minutes
1. Create an agent
From your dashboard, click + New agent. Give it a name — something descriptive like Support Bot or Sales Assistant. The name shows up in the chat widget header by default, so pick something your users will recognize.
New agents start at the setup wizard, which walks you through naming, picking a starter prompt, and adding your first source. You can also skip the wizard and configure everything from the sidebar.
2. Add a knowledge source
Open the agent and head to Sources in the sidebar. Pick the source type that matches what you have:
Pick the source type
Data sources has the full menu — files, websites, text snippets, Q&A pairs, and Notion. The starter pick is usually
Files(drag in a PDF) orWebsite(point us at your help center).Drag your content in
Files accepts PDFs up to 10 MB. Website crawls the URL you give it and lets you cherry-pick which pages to include. Text snippet is just a textarea — paste, save, done.
Watch the storage bar
The right rail shows your storage usage with colour-coded warnings at 70% and 90%. Free tier has a tight cap; paid tiers get more room. Sources you toggle off don’t count.
3. Train the agent
Click Retrain in the right rail. Training reads every active source, splits it into searchable chunks, and stores embeddings the agent uses at chat time. The status pill in the sidebar flips to Training… while it works — usually under a minute for a couple of PDFs.
Failed pages show up in a banner at the top of the Sources page with a Retry failed button. Almost always a transient fetch error — one retry usually clears it.
4. Preview in the playground
Click Playground in the sidebar. The right pane shows your live widget; the left sidebar holds the controls. Type a question your source should answer and check the response.
What to try first
- A question the source clearly answers (sanity check)
- A question the source doesn't cover (does the agent admit it?)
- A loaded question (does it stay on-topic?)
If the answer’s off, tweak the system prompt in the sidebar (it appears in the Instructions field), or add more sources. See Best Practices for prompt patterns that consistently produce better answers.
5. Deploy the widget
Open Deploy in the sidebar — that’s the All channels page. Two featured channels are ready out of the box: the floating Chat widget for your site, and a ChatGPT-style Help page deployed at /help.
Toggle the channel on
The Chat widget card has a switch in its header. Flip it on; the preview updates to your agent’s name and welcome message.
Copy the snippet
Click
Manageon the widget card and grab the embed snippet. Paste it before the closing</body>tag of your site. That’s the entire integration.Verify on the live site
Reload your page. The chat bubble appears in the corner. Click it, ask a question, confirm the response. You’re live.
Next steps
You have a working agent. Here’s what to look at once it’s live:
- Watch real conversations on the Activity page — see what users actually ask
- Skim Analytics for trends in topics, sentiment, and message volume
- Set up Actions so the agent can do things (book calls, capture leads, search the web)
- Tighten the system prompt and add Q&A snippets for any rough answers you spot
Related
Best Practices
Prompt patterns, source curation, and retraining cadence that keep agents accurate over time.
Playground
Tour the model picker, system prompt presets, and live widget preview.
Data sources
Every source type Chatmount supports — what it’s for, what it ingests.
Deploy
Channels, embed snippets, and the Help page surface in detail.