Project Assistant
The assistant is the first thing you see when you open a project. Ask it anything about that project in plain language — it reads your data with the same tools the MCP server exposes, and it can make changes when you ask it to.
It replaces the old overview dashboard. That dashboard answered a fixed set of questions well and everything else not at all; the assistant answers whatever you ask, and can act on the answer.
What it can see
The assistant is scoped to one project and reads only that project's data:
| Area | Examples |
|---|---|
| Web analytics | Visitors, pageviews, sessions, bounce rate, breakdowns by page/referrer/country/device, period comparisons, retention, realtime |
| Events | Custom event stats, timelines, and breakdowns |
| Users & sessions | User profiles and activity, session detail and event streams, frustration signals |
| Funnels & journeys | Funnel conversion analysis, path analysis |
| AI visibility | Presence over time, share of voice, competitors, prompts, topics, cited domains |
| Bots & crawlers | Top bots, timelines, countries, most-crawled paths |
| Backlinks | Detected links, opportunities, reclamation candidates |
| Site audit | Latest and historical audits, their issues, and every page a crawl reached |
| Content recommendations | Open recommendations and their status |
| Webhooks | Configured webhooks and their events |
Some questions to start with:
- "How did traffic change last week compared to the week before?"
- "Which pages lost AI visibility this month?"
- "Which of my pages have the worst audit scores?"
- "Which competitors are cited more than us, and on which prompts?"
What it can do
The assistant can also change things — add an AI-visibility brand, prompt, or topic; start a site audit; update a backlink or recommendation status; create or update a webhook.
It can never do more than you can. Every action is checked against your own permissions at the moment it runs, so a viewer asking the assistant to delete a brand gets the same refusal they would get from the UI.
Team and organization management — inviting or removing members, changing roles, editing organization settings — is deliberately outside its reach, along with creating projects. Those stay where you have to navigate to them on purpose.
Actions that need your approval
Anything that removes data or breaks collection stops and asks first:
- Deleting an AI-visibility brand, prompt, or topic
- Deleting a webhook
- Cancelling a run
- Changing project settings, including the tracked domain
You see exactly what the assistant is about to do — the real arguments, not a paraphrase — and choose Approve or Decline. Declining is not a dead end: the assistant is told you said no and can suggest something else.
Everything else runs without interrupting you: reads are always safe, and creates and updates are recoverable.
Attaching files
Attach a file with the paperclip, by dragging it onto the composer, or by pasting a screenshot straight in.
| Type | Limit |
|---|---|
| Images (PNG, JPEG, GIF, WebP) | 5 MB each |
| 10 MB each | |
| Text, CSV, JSON, Markdown | 10 MB each |
Up to four files per message.
Attachments are not stored
A file is sent to the model for that one message and then discarded — only a note that you attached it is kept in the conversation. A later question cannot re-examine an earlier attachment, so attach it again if you need to ask about it a second time.
Conversations
Each conversation is private to you. Two people looking at the same project have their own separate histories.
- New conversation starts a fresh one and archives the current thread.
- Conversations opens your past threads, titled by their opening message.
Conversations survive a reload — including one paused on an approval prompt, which is still waiting when you come back.
A very long conversation eventually loses its beginning: the assistant reads back dozens of exchanges, but not forever, and the detail of what a lookup returned fades before the questions do. Ask it to remember anything that should outlast the thread — a remembered fact comes back on every future question about the project, here and in Slack alike.
Cost
Assistant usage is metered per project and counts toward your AI spend, the same as AI visibility runs. Reads are cheap; the cost is dominated by how much data a question pulls in.
Availability
The assistant needs an Anthropic API key configured on the server. On Kitbase Cloud it is always available. If you are self-hosting, set ANTHROPIC_API_KEY in the server environment — without it the composer explains that the assistant is not configured and the rest of the dashboard is unaffected.
Next steps
- Assistant in Slack — the same assistant, from a DM or a channel.
- Dashboard Guide — the other views in the dashboard.
- MCP Server — the same tools, from Claude or any other MCP client.
- AI Visibility — the brand-mention data the assistant reads.
- Analytics — how the metrics it quotes are computed.