Approvals
How approval rules make risky actions by people, AI agents, workflows, MCP clients and API keys wait for a human decision, and how approvers decide, escalate and expire requests.
Inbox
How the Inbox shows customer conversations from WhatsApp and web channels, how staff reply with attachments and AI suggestions, take over from the AI, approve AI order proposals and track delivery.
Approvals
How approval rules make risky actions by people, AI agents, workflows, MCP clients and API keys wait for a human decision, and how approvers decide, escalate and expire requests.
Inbox
How the Inbox shows customer conversations from WhatsApp and web channels, how staff reply with attachments and AI suggestions, take over from the AI, approve AI order proposals and track delivery.
How the CentraKit AI assistant works in the Chats page and side panel, including Ask, Plan and Agent modes, models, attachments, commands, approvals, agent runs, sharing, contextual actions, knowledge retrieval and organization AI settings.
Chat is the AI assistant of CentraKit. It answers questions about the organization, plans work and carries out actions with the same permissions and approvals as the person using it. A prompt can never grant access that the user does not already have.
The assistant is available in two places:
| Route | Purpose |
|---|---|
/{orgSlug}/chats | Chat history and a new chat. |
/{orgSlug}/chats/{threadId} | One chat. |
/{orgSlug}/chats/activity | Opens Notifications filtered to items that need action, such as AI approvals and proposed artifacts. |
/{orgSlug}/ai/settings | Organization AI settings. |
/portal/chats/{token} | Read-only shared chat, opened from a share link. |
A chat is a thread of messages. Chats are private to their owner by default. The history groups chats into Pinned and recent periods and can be searched. The actions menu on a chat offers Share, Rename, Download (a transcript), Mute notifications, Pin or Unpin, Archive and Delete.
| Mode | Use it to |
|---|---|
| Ask | Get answers about the organization from its data and knowledge. |
| Plan | Get a structured plan. Review the plan before moving into Agent mode. |
| Agent | Let the assistant carry out the work with tools, as a durable agent run. |
The highest mode an organization can use and the risk level that needs approval are set by platform administrators. See Admin AI settings.
Auto chooses the model, reasoning effort and context window for each request. You can also pick one of the enabled models, a reasoning effort, a context window and a speed (Normal or Fast). A chosen model must support the attached files. When it does not, CentraKit asks you to choose Auto or a compatible model instead of switching silently. Each settled answer shows the model that actually produced it.
The assistant works through tools for customers, assets, tasks, quotes, invoices, orders, purchase orders, promotions, agenda, files, knowledge, inbox, time off, HR approvals, analytics, workflows, integrations, notifications, access, billing and usage, plus web search. Every tool checks the user's permission.
Routine reads run without asking and are summarized as gathered context. Changes show an approval card with Approve and Deny. When several changes are proposed at once, they appear as one checklist, and you can approve all, deny all or approve a subset. Only approved tools run.
An Agent mode request runs as an agent run with steps for planning, tasks, tool calls, approvals, checks and handoff. A run can be queued, planning, awaiting approval, running, checking, retrying, succeeded, partially succeeded, failed, canceled, stopped or escalated. If a connection drops, the chat reconnects to the run. Request human handoff asks a person to take over and records why. Approvals are always decided by people, never by an agent.
Many pages offer AI actions for what is on screen, such as summarizing a customer, reviewing a quote before sending, drafting an inbox reply or explaining an analytics report. Drafts are applied to the form as reviewable changes, and nothing is saved until you save the form. Larger results are stored as artifacts, such as plans, reviews, summaries, customer follow-ups and workflow suggestions. An artifact is proposed, applying, applied, rejected or stale when the page it was made for has changed.
Select New chat, or open the side panel. A new chat offers three starter suggestions from your organization context. Type a message and send it. After each answer, the assistant suggests follow-up questions, and Regenerate suggestions fetches three new ones.
@ or # to reference users, teams and records such as customers, tasks, quotes and invoices. References appear as chips with a preview on hover./web to search the public web./workflow and pick a published organization workflow to run.Select Add attachment to upload files, take a picture or take a screenshot. Images, audio and PDFs are checked against the chosen model. Select Voice input to dictate a message, which is transcribed into the composer.
You can keep typing while the assistant answers. Extra messages wait under Queued messages, where you can edit or delete them. Stop ends the current answer. When a response fails, the notice explains the cause, such as a rate limit, the organization usage limit, an unsupported attachment or an unavailable model, and offers the matching retry or fix.
Mark an answer as Good response or Poor response. Ratings are private to you and store no message text.
Select Share in the chat's actions menu:
/portal/chats/{token} without signing in.A revoked, expired or invalid link shows "Shared chat unavailable".
The assistant searches the organization's uploaded knowledge, such as manuals, policies and handbooks, and can list and read the knowledge files and their versions. Knowledge search is for documents, not for app records like customers or invoices. Organizations upload knowledge under AI, Knowledge in the sidebar, and platform administrators manage platform knowledge in Admin knowledge. MCP clients use the same retrieval through the retrieveKnowledge tool.
/{orgSlug}/ai/settings controls AI for one organization:
| Setting | Meaning |
|---|---|
| Inference region | EU or US pins inference and provider storage to that zone. Global routes for availability and latency with no residency guarantee. Pinning a region can cost more. |
| Chat instructions | Optional terminology and tone for Ask, Plan and Agent. They cannot override permissions, approvals, modes, tools or platform safety rules. |
| Background actions | Whether AI actions may run in the background. |
| Private memory | Whether the assistant keeps private memory per user, and for how many days. |
| External integrations | Whether AI actions may use connected integrations. |
| Daily credit budget | A daily limit on AI credits for the organization. |
| Contextual actions | Every contextual action per page, each with its own on and off switch. |
| Permission | Allows |
|---|---|
organization.chats.view | Open Chats and the side panel and read chats you can access. |
organization.chats.create | Start chats, send messages and start specialists. |
organization.chats.update | Rename, pin and archive chats. |
organization.chats.delete | Delete chats. |
organization.knowledge.view | Let the assistant search organization knowledge. |
organization.general.update | Edit organization AI settings. |
Only the owner of a chat can change, share or delete it and manage its participants. Participants can read and post in the chat. Each tool also needs the permission of the feature it touches. See Access control and RBAC.
retrieveKnowledge. See MCP and AI agents.The chat transport lives under /api/v1/ai/:
| Resource | Routes |
|---|---|
| Chat stream | POST ai/chat, GET ai/chat/{threadId}/stream to resume, POST ai/chat/{threadId}/stop |
| Threads | GET ai/chat/threads, PATCH and DELETE ai/chat/{threadId} |
| Participants and sharing | ai/chat/{threadId}/participants, ai/chat/{threadId}/share |
| Attachments | POST ai/chat/attachments, GET ai/chat/attachments/{attachmentId} |
| Suggestions | POST ai/chat/starter-suggestions, POST ai/chat/follow-ups |
| Voice | POST ai/transcribe |
| Specialists | GET and POST ai/specialists |
Organization routes live under /api/v1/organizations/{organizationId}/ai/: threads, threads/{threadId}/messages, activity, models, agent-runs/{runId} with events, approvals/{approvalId} and handoff, contextual-actions/{actionId} and contextual-artifacts. Shared chats are read from GET /api/v1/public/chats/{token}, and knowledge from POST /api/v1/organizations/{organizationId}/knowledge/retrieve.
Last updated on
Approvals
How approval rules make risky actions by people, AI agents, workflows, MCP clients and API keys wait for a human decision, and how approvers decide, escalate and expire requests.
Inbox
How the Inbox shows customer conversations from WhatsApp and web channels, how staff reply with attachments and AI suggestions, take over from the AI, approve AI order proposals and track delivery.