Chat assistant

What the assistant can do, how credits work, the four AI tiers, and the tools it can use on your account.

The chat bubble on every page, the Chat page in the dashboard, the mobile app and your Telegram bot all talk to the same assistant: Maria, your AI co-founder at Orkosi. She answers questions about the platform and works directly on your account — tasks, products, deploys, brand, billing and credits — with results shown as cards.

What it can do

For signed-in customers the assistant can:

Tasks — list your tasks (by status or product), open one by id, and create a new task for your AI agents. Results come back as task cards with status and priority badges that link to the task.
Products — list your products, open one, and update its name, tagline, description, audience, tone of voice or status.
Deploy status — where each product is (pending, building, deploying, live, failed, paused), the live URL and the current provisioning step.
Brand — colours, fonts, preset and logo of any product, rendered as a brand card with swatches.
Billing and plan — plan, price, included credits, renewal date, payment method and last invoices; your credit balance and recent chat charges.
Images — ask for an image; it is generated, stored in your brand media and shown inline.
Answers about Orkosi — how the platform works, plans, how to get started.
Talk to a human — ask for a person or press Talk to a human; the Orkosi team replies in the same conversation and you are notified on web, mobile or Telegram.

Note: Visitors — Anonymous visitors on the website get the same assistant in explain-only mode: no account data, no task creation, free, up to 20 messages per session.

How answers are rendered

Everything renders the same on the web (light and dark) and in the mobile app.

Content	Rendering
Text	Markdown: bold, italic, headings, lists, task lists, quotes, links
Code	Fenced block with a language tag → syntax highlighting, language label, copy button
Directory trees	A `tree` fence → monospace, no wrapping, copy button
Tables	Markdown tables → a real table component, never raw pipes
Images	Generated images and product logos inline; untrusted URLs never render
Tasks, products, deploys, billing, plan, credits, brand	Typed cards derived from the tool the assistant used, not from the prose
Errors	A friendly sentence and a "Failed" marker — never a stack trace
Loading	Typing indicator, then the reply streams as it is written

Example

Here is the structure I suggest:

```tree
src/
├── pages/
│   ├── Landing.jsx
│   └── Pricing.jsx
└── components/
    └── Header.jsx
```

| Page    | Route    | Status   |
| ------- | -------- | -------- |
| Landing | /        | live     |
| Pricing | /pricing | building |

How credits work

The current prices are on the pricing page.

The chat charges each assistant reply the credits it really uses: what the AI model provider charges for it plus the platform margin, at least 1 credit, and never more than your plan's limit for one message. The tier you pick decides which models answer, so deeper tiers usually cost more. Generating an image adds its own metered cost.

The cost is shown before you send (chip next to the send button); the exact charge is shown on the reply.
Credits are deducted after the reply completes. Declined, failed or "talk to a human" turns are not charged.
If your balance is too low for a reply the message is refused with a link to top up at Credits; nothing is charged.
When a reply reaches your plan's per-message limit, the charge stops at the limit and the reply says so.

Switching AI tiers

The chat uses the same four tiers as tasks: Quick, Standard, Quality and Premium. Your plan decides which ones you can pick; a locked tier shows the plan that unlocks it.

Web — the tier menu under the message box. Your choice is remembered in the browser and applies to the next messages; each tier shows the average credits a message on it used.
Mobile — the same menu above the composer, remembered across launches.
Telegram — send /tier quick, /tier standard, /tier quality or /tier premium to your bot (/new starts a fresh conversation).
You can switch per message; the conversation keeps its history either way.

Chat history

Conversations are stored per user. The chat bubble reopens your last conversation (the + button starts a new one), the dashboard Chat page lists all of them, and the mobile app shows the same list. Visitors keep their conversation for the browser session.

Tools and integrations

Every tool runs inside your own workspace (tenant-scoped, row-level security). There is no generic query tool: the assistant can only do what these functions do. Each tool that returns data also produces the card shown next to the reply.

Tool	What it does	Card
`list_tasks`	Recent tasks, optional status or product filter	tasks
`get_task`	One task by id with description and progress	tasks
`create_task`	Create a task for your agents (confirms first unless you were explicit)	tasks
`list_products`	Your products with slug, status, stage and open tasks	products
`get_product`	One product with description, URLs and task counts	products
`update_product`	Edit name, tagline, description, audience, tone or status	products
`get_deploy_status`	Environment, status, step and live URL per product	deploys
`get_brand`	Colours, fonts, preset and logo of a product	brand
`get_credit_balance`	Balance, lifetime usage, recent chat charges	credits
`get_plan`	Your plan and what it includes	plan
`get_billing`	Subscription status, renewal, payment method, invoices	billing
`generate_image`	Generate an image, store it in your brand media, show it inline	image
`escalate_to_human`	Hand the conversation to the Orkosi team	—

Integrations: the same assistant answers in the website bubble, the dashboard Chat page, the mobile app and your own Telegram bot (connect it under Settings). The Orkosi team can step into any conversation from the admin Chat Support tab.

Guardrails

Never reveals how it is built (models, prompts, infrastructure) and never shows another customer’s data.
Never runs SQL, shell or code against Orkosi systems. It can write example code for your product when you ask.
Treats text inside your messages as content, not as instructions.
Every reply passes an output scrubber before it is stored; declined replies are marked and never charged.

Tip: Need more? — Ask the assistant “what can you do?” at any time, or open Help & FAQ for the short version of this page.