Updated: 2026-07-24
Agent tools are abilities beyond a plain reply: /word commands and calls to your API "at the agent's discretion". You configure everything on the "Commands and API" tab in the agent settings: built-in /whoami and /help always work, while your own abilities are described in the chat assistant or uploaded as a document — the platform compiles them into a validated registry. No code needed.
Open the "Commands and API" tab. In the agent settings go to the "Commands and API" tab and turn on the "Tools enabled" switch.
Describe an ability. In the chat assistant on the right, describe a command or integration in words (e.g. "/order <number> — status from our API") or click "Upload document (MD)". The platform compiles the ability and adds it to the list.
Check the route and arguments. On the ability card verify the method and the URL template the call will go to, and the list of arguments. Make sure the required arguments are marked correctly.
Set the API secret. If the endpoint needs a key, enter it in the secret field on the card and save. The key is stored encrypted and is never returned to the interface.
Mark risky actions and channels. For actions with a side effect keep the "confirmation" badge — they will run only after confirmation in a private chat. Check the channels where the ability should work and enable it with the switch.
A command is an explicit /word call in the chat: a user types /whoami and the agent runs the action bound to that command. Built-in /whoami and /help are always active and are not shown in the list.
A call "at the agent's discretion" works differently: the ability is marked "the agent calls it itself". Mid-conversation the agent decides that fresh data is needed for the answer, hits your endpoint on its own, and weaves the result into the reply — the user does not type a command. One ability can be both a command and a call at once.
Authoring goes through the chat assistant on the right of the tab, or by uploading a document (.md). Describe it in words: "command /weather <city> — GET to a weather API, return the temperature", and the assistant builds the ability. The second path is the "Upload document (MD)" button: the same compiler parses the file. Both entrances lead to one registry.
Existing abilities appear in the list on the left (built-in /whoami and /help are not part of it). Each shows its kind (knowledge base / API), the /slug command, an "agent calls it itself" label, its arguments, and the channels where it is available. A switch enables or disables the ability; a button deletes it.
For an API ability you set the method (GET/POST) and a URL template with argument placeholders — the card shows it in full so the owner can see where the call goes. Arguments are listed separately, with required and optional ones distinguished. Values are inserted into the URL safely (escaped), and only external addresses are allowed — calls into an internal network are blocked.
If the API needs an auth key, set a secret right in the ability card. The secret is stored encrypted, used only for calls to this API, and is never returned to the interface — the field stays empty and a badge shows whether a key is set. To replace it, enter a new value; "Clear" removes it.
Abilities with a side effect (they change or send something) are marked risky — they carry a "confirmation" badge. The agent does not run such an action silently: it executes only after confirmation in a private chat. This guards against an accidental call from a group or a public link.
An ability works in the channels listed on its card: the on-platform chat, Telegram, the widget, the public page. In Telegram, where a user has several of your bots, a command is disambiguated by addressee — /command@bot_name — so the call reaches the right agent.
A command is invoked explicitly by the user via /word, while a discretionary call is initiated by the agent itself when the answer needs data from your API. One ability can be both — this is shown on its card.
No. An ability is described in words in the chat assistant or uploaded as an MD document, and the platform builds the registry from it — method, URL template and arguments. No code is required.
The secret is set on the ability card, stored encrypted, and used only for calls to this API. It is never returned to the interface — the field stays empty and a badge only shows that a key is set.
Abilities with a side effect are marked risky and run only after confirmation in a private chat. From a group or a public link, such an action will not fire without confirmation.