Console assistant
The Console assistant helps you work on bots. It is on every page of the Console. Choose Assistant in the top bar. A panel opens at the side.
It is not a bot of your organization, and it is not the Personal bot. GEKI provides it, and you cannot change its instructions.
Who sees it
Owners, Admins, and Builders see the Assistant button and the panel. The assistant works with your rights, cut down. It is not available when GEKI has turned off internet for the bots of your organization.
The panel
On a wide screen the panel pushes the page aside, like the navigation. On a narrower screen it lies over the page. It stays open when you go to another page, and the conversation continues.
Under the title, a small line with an eye icon says which page the assistant sees, for example "Context: Bot GEKI Slides · Editor". The panel opens with a short text about what the assistant does. It shows no starter questions. When the assistant answers with a link to a Console page, the link opens in the Console and the panel stays open. Other links open in a new tab.
The conversation
There is one conversation for you in each organization. The assistant continues it while you were active in the last 30 minutes. After that, it starts a new conversation. New conversation in the header of the panel starts one at once. The old conversation is not deleted. You find it under Sessions, where it is named Assistant.
You can attach files in the message box. The assistant can read them in its workspace.
The first conversation
On Start, Get started begins with your first company bot. Describe what it should do, or insert the example, and select Build it. The assistant panel opens with your description. The assistant builds a first, small version right away. It does what you described and nothing more. When the bot works with your company knowledge, it gets an empty Knowledge Base.
With a Knowledge Base, the assistant reads the start page and the imprint of your website. Then it asks what goes into the Knowledge Base first: your company master data, a specific area you name or a document, or nothing for now. When your website shows no company data, it says so and offers only the area or document and nothing for now. It writes only what you chose. Then it tests the bot with one question and proposes publishing it.
When the bot needs an account that it cannot reach yet, such as a mailbox, the assistant builds it without and tells you what is missing. When the bot is live, it tells you that the next step is your Chat address on Start, and links the bot in Chat once you saved the address. The assistant answers in your language.
Set up a bot with the assistant
On the Bots page, New bot asks for a name and a short description. Then choose:
- Set up with the assistant. The Console creates the draft, opens its editor, and opens the panel with a first message. The assistant asks what it needs to know and writes the instructions.
- Set up myself. The Console creates the draft and opens its editor.
What it sees
The assistant knows which page you have open and what is selected on it. When a sheet or dialog with a form is open, it also sees the form and what you typed so far, for example the name and the URL of a new MCP connector. It never sees a password or a secret you typed. It only learns that the field exists. It also reads the error messages that the page shows. Ask "Why does this not save?" and it answers from the message and the form.
It knows the pages of the Console that your role can open and links to them. It does not link to a page you cannot open.
Ask the assistant buttons
Some places have an Ask the assistant button: a failed run, the warnings of a bot, and a failed publish. The button opens the panel and sends a short message in your name, for example "Why did run 123 fail?". It starts a new conversation.
What it can do
Ask it, for example, what a bot does, to improve the instructions, or why the last run failed. It can:
- Read bots, runs, sessions, Knowledge Bases, and usage that you may read.
- Change the draft of a bot, and create a new bot as a draft.
- Create a skill, or a new version of a skill. It never overwrites one.
- Create and change Knowledge Bases, write their pages, and attach them to a bot draft.
- Search the web, unless an Owner or Admin turned web search off under Settings, Capabilities, Assistant.
- Add a trigger, or change one. A trigger it saves is turned off.
- Create an empty vault and attach it to a bot draft. It never enters a credential.
- Check an MCP server, and list the tools of an MCP connector.
- Read this handbook, to explain a feature or build something new. It tells you when GEKI cannot do what you ask, for example that there is no website widget yet, and suggests the closest option.
- Test a bot in a preview and read the answer.
- Upload a file to attach it to a bot draft.
- Ask you a question and wait for your answer.
After each answer the Console loads the bot again. When the editor is open, it asks you to reload.
What it proposes
It changes drafts only. Anything that goes live or cannot be undone, it proposes. The proposal shows in the answer as a card with a button. The card has a short text from the assistant and a line from the Console that says what happens and to which object, for example "Pause · Bot Support" or "Delete · Skill pdf-tools". If the two do not match, trust the line of the Console. You can open the details of the call on the card. Nothing happens until you click.
Typical proposals are to publish, pause, resume, or roll back a bot, to delete a bot, a skill, a trigger, a Knowledge Base, or an MCP connector, to create or run a trigger, and to create an MCP connector.
A click runs the call with your own rights, like the buttons on the pages. A delete has a red button and asks you to confirm once more. When a bot has new capabilities, the card opens the same list as the Publish button, and you confirm each item. If it works, the card shows that it is done and cannot run twice. The Console then posts a short message as you, for example "Done: Pause · Bot Support", so the assistant knows. If it fails, the card shows why, and the Console posts "Failed:" with the reason. The card is disabled when the object is gone or when your role may not do it. You can ignore a card. The assistant does not wait for it.
What it cannot do
It never runs these actions itself.
- It cannot change who can use a bot, pin a bot, or change its status.
- It cannot change or delete vaults, and it cannot change provider connectors, MCP connectors, OAuth apps, model settings, members, or groups on its own.
- It cannot enter a credential, a token, or an API key. It links the page where you do it.
- It cannot start a chat with any bot other than a preview.
A card never changes credentials, API keys, members, invites, groups, roles, billing, credits, packages, session takeovers, or the organization itself, whatever your role. For these, the card has no button. It links to the Console page where you do it yourself.
Its changes show in the editor as unpublished changes. Review them, then publish the bot yourself or with a card.
How it works
For each answer, GEKI gives the assistant a key that holds your rights minus the ones above. The key ends with the answer. The assistant runs in a workspace with limited internet access. It reaches only the GEKI API, this handbook, the package sites it needs, and your own sites: the domain of your company email, the website of the publisher under Settings, and the extra domains that an Owner or Admin lists under Settings, Capabilities, Assistant. A new conversation uses the list as it is then. Add up to 20 host names, one per line, for example docs.example.com. Its work counts as your usage.
An Owner or Admin picks the model of the assistant under Models, Default models, Assistant model and Assistant thinking. Until they choose, the assistant uses the GEKI default, DeepSeek V4.1 Flash with High thinking. Reset goes back to it.