Concepts

Grants and approvals

You say yes at install time. One gate checks every action. Approvals let you edit the details. Chat can get session grants. Runs you are not there for can be approved ahead.

Nothing in Recued touches your accounts unless you installed the pack and said yes ahead of time. That yes is called a grant. You give the grant before anything runs. The gate checks it at the moment of the call. Both steps leave a record.

#Grants are made at install

Installing a pack is the moment you decide what it may do. Before anything is installed, Recued shows you:

  • every recipe the pack will install;
  • the account (the connection) it can use, if any;
  • the access level it asks for, such as read, or read and write;
  • the audience, which means who may use those grants. Unless you change it, that is only you;
  • whether it may see message bodies or use local tools, if it asks for that.

The grant sticks to the installed pack. One recipe cannot quietly make it bigger. You can change it later without installing anything again.

#One enforcement boundary

Every action goes through the same gate before it runs. It does not matter if you clicked Run, a timer went off, or an outside AI asked:

text
request
  ↓ resolve who is asking (actor) and on which channel
  ↓ resolve the installed capability and its contract
  ↓ check the grant and the operation's risk tier
  ↓ execute locally
  ↓ write the audit record

There is no special path around it. The server's own timer, your chat messages, and a connected AI all meet the same checks.

#Risk tiers and approvals

Reads run when you granted them. Writes can be set to ask you first. Sending mail or a Slack message asks first, even when you are the one asking. It keeps asking until you decide otherwise. An action that cannot be undone, such as most deletes, asks every time. Other writes follow their own setting. In your own chat, many run at once. An update to a HubSpot or Salesforce record is one. To make an action ask every time, open its pack's Permissions tab and set its approval to Owner · Always.

An approval is a pause. The request reaches your paired devices. It is saved and waits until you answer. Nothing moves until you say yes or no. The request may carry details, like an email draft or a record change. It names each record by the name you know it by, such as Family car or Mia Whitfield (HubSpot Contact), not by an id. It names each of your contacts the action is aimed at, such as Mia Whitfield. That covers an email's recipients, an invite's guests, a task's assignee and the number a text goes to. An email sent in its encoded form shows whom it goes to and its subject. An approval sent to Telegram, Slack or your email lists these details too. You can edit the details before you approve. Sometimes one source sends a burst of the same action many times. Then Recued asks once for the whole batch. It does not ask once per item.

#Session grants

When you work in chat, you can choose to relax the asking for the rest of that session:

  • Exact means approve this one exact action, once.
  • Batch means approve this shape of action for the burst it belongs to.
  • Open means allow this kind of action, within one area you set, for the session.

An approval that offers Allow this session says what it lets happen. It names the recipe that may go ahead without asking, how many more times, and for how long. It lists what must stay the same on each later run. For an email, that is who it goes to and the mailbox it is sent from. For a notice to a booking's visitor, it is the booking. It also names what may change, such as the subject and the text. A change to anything on the list asks you again. Some actions allow only the same thing again, with every detail the same, and the approval says that instead. The approval email and the Approvals page show this. The approvals pop-up in the top bar leaves it out to stay short.

Some changes always need your own look, so they never offer Allow this session. These are a proposed commitment, a quote, a job a customer accepted, and a customer payment.

If you approve the same thing many times, Recued may suggest a standing rule. Only you can make one. A suggestion never switches itself on.

#Don't ask about browser actions

Your AI can work in your browser for you. Below the chat's box is a switch: Don't ask about browser actions. Turn it on, and the job you ask for goes ahead without asking at each step. That covers opening pages, typing, clicks and sent forms. It covers every site the job goes to.

The switch stays on while the job goes on. That includes your AI's questions to you and your answers. It ends when you turn it off. It also ends when you send a message with it off, or after 8 hours.

Some things still stop or ask:

  • Sites you blocked are still refused.
  • A sign-in code is still asked for at its step.
  • Putting one of your files into a page still asks.
  • A change outside your browser, such as an email or a record, still asks.

#Approve ahead

Some runs happen while you are away. A schedule starts one. So can a webhook, an event trigger, an auto-run you switched on, a public form, or a scheduled prompt. A change such a run makes waits for your approval. You can approve it ahead instead.

There are two ways:

  1. Open the trigger's own page, and open What it may do without asking. Tick the changes to allow, and how many times a run. Then choose Allow without asking.
  2. Or answer a waiting approval with Always allow for this webhook. The last word names the trigger: schedule, trigger, auto-run, form or scheduled prompt.

A webhook's page is its card under Connections. A schedule's page is under Automation. A form's page is under Reception.

What you allow stays on until you end it. End it on the trigger's page, or under Contracts. Contracts lists it under Happening without asking. A run's page says what it did without asking.

#What still asks when you approve ahead

A yes given ahead is narrow on purpose:

  • It covers only the changes it names.
  • It covers only that recipe, as it is now. A change to the recipe asks you again. So does a change to a scheduled prompt's words.
  • A change that deletes still asks.
  • A change you set to ask every time still asks.
  • Some changes always need your own look. A proposed commitment, a quote, and a job a customer accepted still ask.

A scheduled prompt's AI decides what it writes, and its approval says so. Its browser actions can be approved ahead too, on any site. That allows up to 100 a run. Sites you blocked are still refused. Putting your files into a page still asks, and so does a sign-in code.

#When something is turned off for you

A switch is an on or off for one thing on the Contracts page. Your own row there is Self (you). Is a switch off for you? Then a call that needs it does not run. Recued tells you what is off and where to turn it on. For example: "'mail.search' is turned off for you, so it did not run. To use it, turn it on under Self (you) → Ops on the Contracts page."

In chat, the fix shows under the answer. Press Turn it on, then ask again. The run dialog has Turn it on too, for a program that is switched off for you.

An AI app, or anyone else you gave a contract, is told less. It reads only that the thing is not available for that contract. It never sees your switches.

#Doors

A door is a way in from outside that you open on purpose. MCP is a door for AI agents. Reception is a door for public pages and requests coming in. Messenger is a door for chat platforms. Each door carries its own grants, tool by tool. Its secret key comes from the door itself. So when you close the door, the access it gave is gone too.

#When something goes wrong mid-flight

Sometimes a gated call crashes after the decision but before the answer comes back. Recued does not quietly try again. It parks the action as "in doubt". Then it asks you what happened before anything goes on.

Recued Docs

Search documentation

Start typing to search the documentation.