Kit, the AI agent
Ask Kit to do anything you can do in GraphKit, from drafting replies to building and launching campaigns. Kit acts with your permissions and asks before anything consequential.
Kit is the assistant built into GraphKit. It works across your sending accounts, leads, lists, campaigns, inbox and analytics, using the same actions you have in the app.
Kit lives at the right end of the top bar on every page. Click its face or press ⌘J (Ctrl+J on Windows) to open the panel. In ⌘K (Ctrl+K), type a question and pick Ask Kit to send it straight to Kit. The full page at Kit keeps your conversation history.
What the top bar tells you
The face next to the page's actions shows what Kit is doing, and a short status appears beside it only when there is something to say:
- Idle: just the face, blinking now and then.
- Working: what Kit is doing right now, such as "Reading 14 new replies" or "Enrolling 32 leads".
- Waiting for you: "2 to approve" when Kit has prepared actions that need your decision.
- Done: a short smile when a run finishes, then a dot until you open the panel.
- Autopilot on: the current activity or the next check ("Next check in 6 min"), see Autopilot.
You can dock the panel while Kit works: it shrinks to a small bar at the bottom right and keeps going. Inside the panel, a status line shows the current step and the elapsed time, a trace lists every step (open any step to see its result), and each run ends with What Kit did, with links to what it changed.
What Kit can do
| Area | Examples |
|---|---|
| Inbox | "Draft replies to unread LinkedIn messages", summarise a thread, label, assign, add notes, mark as read, send a reply (after you approve it) |
| Campaigns | Create a campaign from a template or from scratch, write and validate the sequence, attach accounts and lists, simulate it, launch, pause and resume (launch and resume ask first) |
| Leads | Search and filter, create and update leads, build and combine lists, import a CSV or XLSX you attach (Kit proposes the column mapping and shows duplicates first), add to suppression |
| Accounts | Check status and limits, pause and resume, give you the link to connect or reconnect an account |
| Analytics | Answer questions with your real numbers, export a CSV you can download |
| Research | Read a public web page about a prospect or company, and search the web when your server has search set up |
| Team, developers, billing | List people and API keys, invite a teammate, create an API key or webhook, prepare a checkout link (all ask first) |
Kit never connects an account for you and never asks for passwords or codes: it gives you the link and you finish it yourself.
Kit acts as you
Every action runs with your role and permissions in the workspace, through the same checks as the app. If you are a viewer, Kit can read and answer questions but cannot change anything. If something needs a permission you don't have, Kit says so instead of looking for a way around it.
Everything Kit does is recorded in the workspace Audit log as "Kit on behalf of" you.
Approvals
Kit pauses and shows you exactly what will happen before any consequential action:
- sending a message (you see the final text, with variables filled in, and can edit it)
- launching or resuming a campaign or a sending account
- enrolling more than a handful of leads, confirming an import
- deleting, suppressing or un-suppressing
- inviting people, creating API keys or webhooks
- anything billing
Approve, edit then approve, or reject (optionally with a note Kit reads). When Kit prepares several actions at once you can approve or reject them together. Workspace admins can turn off approval for a few low-risk kinds of action in Settings → Kit (off by default). Messages, deletions, invitations, developer keys and billing always ask.
Attachments
Attach files with the paperclip, drag them onto the composer, or paste a screenshot. Kit reads images (PNG, JPEG, GIF, WebP up to 5 MB), PDFs, CSV, XLSX, text, Markdown and JSON up to 20 MB each. Large spreadsheets are read in pages; Kit tells you when it can't see a whole file at once, and never cuts it off silently. Files are stored privately and only you (and the conversation they belong to) can open them. Kit can also give you files back, such as a CSV export.
Memory
In Settings → Kit → Memory, write down what Kit should always know: your ideal customer, tone of voice, offer, and do and don't rules. Kit reads these before writing copy or replies. Ask Kit to "remember" something and it saves a note there too, where you can edit or delete it.
Kit also knows which page you are on, so "this campaign" or "this list" means the one you are looking at.
Model
Kit runs on Claude Haiku 5.5. There is nothing to choose or configure.
Bring your own key
Organization owners can add their own OpenRouter or Anthropic key in Settings → Kit. Keys are checked with the provider, encrypted, and never shown again. Usage on your own key is billed by that provider and does not count against your GraphKit AI allowance, or Autopilot's.
AI allowance
Each organization gets a monthly AI allowance for Kit on GraphKit's model keys, per paid seat. Settings → Kit shows what you've used this month. When it runs out, Kit says so; it resets on the 1st, and an owner can add your own key to keep going.
Safety
- Kit respects every daily and weekly limit, consent requirement, opt-out and suppression. It does not help with evading platform detection.
- Text from inbox messages, lead profiles, web pages and attachments is treated as data. Kit does not follow instructions it finds there, and consequential actions still need your approval.
- Kit does not invent numbers or results. If a tool returns nothing, it says so.
Use Kit from other AI tools (MCP)
Kit's tools, and everything else you can do in GraphKit, are also available as a remote MCP server for Claude, ChatGPT, Cursor and other MCP clients at https://<your app domain>/mcp. Most clients need only that URL: you sign in and choose the workspace. See Use GraphKit from Claude, ChatGPT and Cursor.
Tools act inside the chosen workspace as you. Consequential tools never run over MCP: they return a pending approval that appears on the Kit page under Approvals, and the client can check it with the get_approval tool.
White label
On an agency's brand, Kit uses the brand's assistant name (set the AI assistant name label in branding) and, when platform branding is hidden, the brand's accent colour. It never mentions GraphKit there.
AI reply classification and drafts
Every reply is sorted into interested, question, referral, out of office, not interested or unsubscribe. Categories drive labels, notifications, campaign steps and analytics, and your team can correct them. AI drafts help you answer faster.
Use GraphKit from Claude, ChatGPT and Cursor
Connect any MCP client to GraphKit with one URL. Your AI assistant can then work with your campaigns, leads, inbox and analytics, as you, with approval before anything consequential.