Core Concepts¶

Six things to know. Here's how they fit together:

flowchart LR
    STORE[Store] -->|task bound to it| TASK[Task]
    WEB[All Stores] -->|task with no store| TASK
    TASK -->|seller-console work is split into| SUB[Sub-tasks]
    SUB --> STORE
    SCHEDULE[Schedule] -->|fires| TASK
    KNOW[Knowledge & Skills] -->|starts with prior experience| TASK
    TASK -->|records what it learned| KNOW
    INTEG[Integrations] -->|email · chat bots · calendar · drive| TASK
    PROFILE[AI Profile] -->|which model| TASK

The task is the centre: the store decides which logged-in browser it can use, a schedule decides when it runs by itself, knowledge & Skills decide how well it knows your shop, integrations decide what it can call besides the browser, and the AI profile decides which model does the work.

Store¶

One store = one logged-in browser + one memory. The login is either a Ziniao browser profile or a dedicated Chrome instance Vibe Seller keeps just for it; the memory is a folder of markdown files.

An agent works inside its own store's browser and never borrows another store's login, and writes that clearly belong to another store get rejected — so Store A's hard-won experience doesn't leak into Store B's notes. Three stores can run at the same time without interfering.

Choosing between the two backends: Which browser backend; which credentials Ziniao needs and where they live: Ziniao: Setup & Common Errors; how parallel stores are arranged: Running several stores at once.

Task¶

You write one sentence; the agent plans the steps, drives the browser itself and hands back a report. Every step and screenshot stays in the task detail and can be replayed.

There are two scopes — the only choice you need to think about up front. Pick a store in the sidebar, or pick the last row, All Stores (subtitled "No Store"):

	Task bound to a store	Task created under All Stores
Which browser	That store's own logged-in browser (Ziniao profile / its own Chrome)	A generic Chrome with no store logged in — public pages only
Good for	Work inside one shop: ad tuning, listings, report exports, stock checks	Cross-store and public-web work: tracking shipments, looking things up, comparing shops
Opening a seller console	Do it directly	Not allowed — it splits the work into sub-tasks, each bound to one store, which use that store's login
Execution mode	Follows your default (a fresh install defaults to Auto)	Always Review plan first — can't be turned off
Where notes are written	That store's own folder	Shared cross-store knowledge, then each sub-task writes back to its own store

The app says it itself: "This is the cross-store task list. These tasks have a general web browser for public pages (search, tracking/logistics) plus email and messaging — but not a specific store's backend."

Auto / Review — the picker in the task footer: Auto means "Skip review, run immediately", Review means "Draft a plan, run after approval". Changing it once also saves it as your default (you can set it up front in Settings → General → Task Defaults → Default Execution Mode). For anything hard to roll back — bid changes, budget moves, bulk listing changes — review first. After reading the plan, hit Confirm & Execute, or Request Changes to keep talking.

A parent task and its sub-tasks form a tree: the parent decides what needs doing and merges the findings, the children do the in-store work. When the parent finishes dispatching it doesn't complete — it parks in Waiting until every child has ended, then closes itself with a combined summary of the results. The default shape is one sub-task per store; different country sites of one store are handled inside that same sub-task rather than split up. Deleting a parent deletes the whole tree.

How many tasks run at once is set in Settings → General → Max Concurrent Agents (default 2). Tasks on the same platform in different countries take turns to avoid clashing; everything else runs in parallel in new tabs.

Day-to-day usage: Tasks.

Schedule¶

Turns a one-off sentence into do this every Monday. A schedule creates ordinary tasks when it fires and keeps their history attached to itself.

A store-bound schedule fires exactly one task. An All Stores schedule picks its shape once, when you create it: Fan out (one task per store) or Single (one task), with an optional Run stores one at a time so no two shops share the browser at the same moment.

For schedules you create yourself, the plan is produced once, at creation: after you save it the agent runs the planning pass and freezes the result on the schedule — there's no approval click — and every later fire copies that plan instead of re-planning ("A plan is frozen on this schedule and will be reused for every fire."). You can read the frozen plan any time; to change it, edit the schedule's text — the plan is then marked "Plan stale" and the schedule won't fire until you hit Re-plan.

What suits a schedule and how to revise a frozen plan: Schedules.

Knowledge & Skills¶

The agent's memory is plain markdown on your disk — readable, editable, copyable between machines — all under ~/.vibe-seller/ (that's the Knowledge Directory line in Settings → AI). The page is called Knowledge and has two modes: Files for editing by hand, AI Chat for telling it in words and letting it file things.

Project Knowledge — the platform knowledge that ships with the release, read-only;
Local Knowledge — experience, rules and tips shared across all stores;
one folder per store — that shop's page mechanics, dead ends it already hit, your own requirements;
Skills — operation manuals written for the agent. Amazon, Noon and friends ship built-ins (read-only); ones you write or import are yours to edit.

Files the agent downloads while working (reports, statements) go under Run Data, filed by store and month, instead of mixing into the knowledge. After a run the agent updates what it learned, so the next one starts smarter. Where the files are and how to write a Skill: Knowledge & Skills.

Integrations¶

Besides driving the browser, the agent can call external services directly: Email (read store mailboxes, draft replies), Chat bots (WeCom / DingTalk / Feishu / Slack), Dida365 / TickTick, Google Workspace, the Amazon Ads API (ad tasks hit the API instead of clicking through the console) — and, pointing the other way, MCP, which lets Claude, Codex and similar tools drive Vibe Seller.

One task can mix them: pull stock in the browser → build a Google Sheet → post the summary to a group chat → drop a todo into TickTick. Everything is hooked up under Settings → Integrations, no code. See Several integrations in one task in the Integrations overview.

AI Profile¶

One AI profile = one provider + one model + one API key, stored on this machine. Use DeepSeek for cheap routine checks and switch to Claude when something needs the big model — you just pick a different profile, nothing else changes. The task footer has a profile picker.

Schedules have no picker of their own: each fire uses whatever your default profile is at that moment, so changing your default changes the model for every schedule that follows the default. The "Sync schedules when default changes" switch under Settings → AI additionally moves schedules that were pinned to a specific profile. Out of the box that default is the built-in Claude profile ("System default environment" — it rides this machine's Claude Code session, so a signed-in subscription works with no key to enter).

Where to get keys and which preset to pick: Configure AI Models.

Ready? Go on to the Quickstart; if something breaks, the FAQ.