Setup Guide
Create a knowledge admin account, connect Notion, Slack, and Google Drive via MCP, verify sources, and deploy your agent. Early access.
On this page
Early access / preview. This guide describes the current setup flow. Screens and connector options may change before general availability.
This guide takes you from zero to a live, grounded agent: create an admin account, connect your knowledge over MCP connectors, verify what got indexed, and deploy the agent to a web widget, Slack, or the API. It pairs with the Agent API reference.
1. Create a knowledge admin account
Go to admin.praktickai.app and sign up with your work email. The admin console is where you manage connectors, knowledge sources, API keys, and guardrails. Your first workspace is created automatically; invite teammates from Settings → Members.
Use a shared team address (e.g. ai-admin@yourcompany.com) as the workspace owner so access survives staff changes.
2. Understand MCP connectors
MCP (Model Context Protocol) is an open standard for connecting AI assistants to external tools and data. Instead of exporting files, you authorize a connector once and PraktickAI reads the content live, respecting the permissions you grant. Connectors are read-only for knowledge: the agent indexes and cites your content but never edits it.
| Connector | What it reads | Authorize with | Scope to grant |
|---|---|---|---|
| Notion | Selected pages & databases | OAuth | Read content of shared pages |
| Slack | Selected public channels | OAuth (bot install) | Read messages in chosen channels |
| Google Drive | Selected folders & docs | OAuth | Read-only access to picked files |
Grant the narrowest scope that works. Share a single Notion teamspace, add the Slack bot only to the channels you want indexed, and pick specific Drive folders rather than the whole account.
3. Connect Notion
In the admin console open Connectors → Notion → Connect and complete the OAuth flow. On Notion's consent screen, select only the pages or databases you want the agent to read. Back in the console they appear as available resources.
Connectors → Notion → Connect → (Notion OAuth) → select pages → Save4. Connect Slack
Open Connectors → Slack → Connect and install the PraktickAI bot into your workspace. Invite the bot to each channel you want indexed with `/invite @PraktickAI`. Only channels the bot is a member of are readable.
/invite @PraktickAI # run this in each channel to index5. Connect Google Drive
Open Connectors → Google Drive → Connect and authorize with your Google account. Use the folder picker to choose specific folders or documents — the connector requests read-only access and never modifies your files.
6. Add and verify sources
For each authorized connector, add the specific resource as a knowledge source under Knowledge → Add source (or via `POST /v1/agent/knowledge/sources/`). Each new source starts in `syncing` and moves to `synced` once indexed.
curl https://api.praktickai.app/v1/agent/knowledge/sources/ \
-H "Authorization: Bearer pk_live_xxxxxxxxxxxxxxxx"Verify the source shows `status: synced` and a non-zero `documents` count, then ask a test question in Knowledge → Playground and confirm the answer cites the expected source.
If a source stays empty, re-check that the bot/OAuth grant actually covers that content — the most common issue is the Slack bot not being invited to the channel, or a Notion page not shared with the integration.
7. Deploy the agent
Once sources are synced, deploy the agent through one or more surfaces:
| Surface | How | Best for |
|---|---|---|
| Web widget | Copy the snippet from Deploy → Web widget into your site | Customer-facing help on your website |
| Slack | Enable Deploy → Slack so teammates can @-mention the agent | Internal Q&A over company knowledge |
| API | Call `POST /v1/agent/messages/` from your backend | Custom apps and existing chat UIs |
<script
src="https://admin.praktickai.app/widget.js"
data-agent-key="pub_xxxxxxxxxxxx"
async
></script>The web widget uses a public key (`pub_...`), not your secret API key. It is safe to embed in the browser and only allows sending messages to your agent.
8. Review answers and set guardrails
Under Guardrails you control how the agent behaves when it is unsure. Recommended defaults for a new deployment:
- Grounded-only: refuse to answer when no source matches, instead of guessing.
- Fallback message: a friendly "I'll hand this to a colleague" when grounded is false.
- Human handoff: route low-confidence email replies to a draft (see the Email API autosend threshold).
- Review log: skim the answer log for the first weeks and thumbs-down anything off, which tunes retrieval.
Start conservative — grounded-only answers plus a high email autosend threshold — then relax the settings as the answer log shows the agent is reliable.
That's it — your agent now answers from your connected knowledge. Continue with the Agent API reference to build a custom experience, or the Email API reference to automate mailbox replies.