Integration & Automation

Use Bricksta in Claude and ChatGPT

Instead of searching in Bricksta, you can ask in the chat: “Show me the running customer onboarding cases” or “What is due for me today?”. If your organisation has enabled it, the assistant also does work for you: “Create a case from the Customer Onboarding template for Acme Ltd” or “On #227, enter ‘Bricksta Analytics’ in the Product field”. To do this, connect Bricksta to Claude or ChatGPT once. The assistant always works with your permissions: it sees and changes only what you may in Bricksta. Without enabling it, it only reads.

On this page

What you can ask

  • Cases of a template, e.g. “Show me the running customer onboardings” — with next step, progress, owner, due date and a link into Bricksta.
  • Cases for a customer, e.g. “Where is the onboarding of Acme Ltd stuck?” — the search covers title and number.
  • Cases by step and field value, e.g. “Which quotes for Elbmarsch Logistik are at ‘Approve quote’?” — the same filters as the case list: the step by name (current phase; completed cases pass the filter as in the app) and a template field with its exact value (case does not matter, several values = any of them). Fields hidden from you or the AI give no match, computed too. In Claude the result appears as a “Cases” card: your filters at the top, below them the cases grouped by current step, each with progress, assignee and due date; a tap opens the case, “Open in Bricksta” opens the case list with the same filters.
  • A single case in detail, e.g. “What is open on #227?” — status, progress, phases, open tasks with assignees and due dates, and the details entered in the case and its tasks (current round only, only fields you may see). Anything a customer entered via the portal, a guest link or a form is marked in the chat as “entered from outside”. Fields hidden from the AI never appear, computed too. In Claude the case also appears as a card: the steps of each workflow with the current step, what it is waiting for, the case deadline and what happened last.
  • Analytics, e.g. “How many cases are overdue?” or “Where are the customer onboardings stuck?” — the same figures as on the dashboard (7, 30 or 90 days), and per template the usual duration and the phases where running cases exceed their norm the most. This requires the permission to view organisation analytics. In Claude the result appears as an “Analytics” card: running, overdue and completed at the top, below it the lead time in plain words and the bottlenecks of the running cases — one below the other, without tabs. With too few completions it says “not assessable” instead of a number.
  • Analytics compared with the period before, e.g. “Are we faster than last quarter?” — below that, the “Analytics” card shows “last 90 days vs. the 90 days before” for lead time and per phase or template, with “better”, “worse” or “same”, exactly as on the dashboard; below five completions “no comparison possible”.
  • The history of a case, e.g. “What happened on #227 recently?” — the same entries as the history in Bricksta, with comments.
  • The attachments of a case, e.g. “Which files are attached to #227?” — name, type, size, task and whether a customer uploaded them via the portal or a guest link. The chat does not read the file contents. Files in fields hidden from you or the AI do not appear, computed too.
  • Finding cases by their details, e.g. “Which cases concern Nordlicht Logistics?” — the fields of the cases and their tasks are searched, only fields you may see. In Claude the result appears as an “At a glance” card: the cases grouped by template, each with status, due date and when something last happened; a tap opens the case in Bricksta. Cases you may not see do not appear and are not counted. If nothing is found, the card suggests a shorter spelling.
  • Datasets, e.g. “Which customers are in the Customers dataset?” — the key field is searched.
  • Team workload: open assigned tasks per person (if your organisation enabled it) or per role. This requires the permission to view organisation analytics.
  • Your open tasks, grouped like “My Work”: overdue, due today, later, unassigned and waiting on others (submitted, the approval is with someone else). In Claude they appear as a card: up to five tasks with the due date spelled out, “Show all” opens the full list, and each row opens the task in Bricksta. The card changes nothing, it only shows.
  • Your organisation's templates.

Reading and changing templates

With the permission “Manage templates” the assistant reads how a template is built — e.g. “Explain how Customer Onboarding is set up” — and, on request, changes its draft.

ReadFirst the outline (settings, workflows, outcomes, webhooks, check findings), then per workflow the phases, tasks and fields; plus the changes since the last version and the same findings as in the editor. Every answer names the version and whether an assistant changed something recently. Secret values are never handed out.
ChangeNew templates (name, title pattern, auto-start); in the draft create, change, deactivate workflows, phases, tasks, fields with choice values, formulas, groups, “Ends when …”, completion emails and message, AI and document steps. Sending waits for review; delete only unpublished items, never mails. Steps with fields: all or nothing. Tasks also: assignee, due date, field assignments, approver, completed on payment, look-back, waiting for the parent case. It adds dataset fields (display column, “Fill other fields”) and relinks a published field’s dataset from the next version. A person publishes.
PathsWhat waits for what (workflows, phases, tasks) and branches such as “if rejected, back to rework”. It rejects loops; published ones it only deactivates.
Time rules“If still open after …”, “wait …” with a date field (which becomes internal only), backjump.
OverviewsAdmins only: add and change list columns, key figures, reference lists, tiles (from the next version); delete unpublished ones.
Not changeTasks for external parties or external approvers, the intake form, file fields, “Only entries matching” and “When the dataset changes” (on such fields not their dataset or type either), targets of return steps in running cases, sending without review or on behalf of others, deleting a completion email, and anything already published except overviews, the dataset link, completion emails and settings (as a proposal; not title pattern or auto-start) — for those it gives you the link into the editor (“only changeable in the editor”).
EnableAn admin switches on “Change template drafts” under Organisation › AI connections (only with “Read and write”); you tick the box of the same name when connecting.

What you can have the assistant change

For this, an admin sets the organisation to “Read and write”, and you tick “May write on your behalf” when connecting (steps 1 and 3). Then you can say, for example:

  • Start a case: “Create a case from the Customer Onboarding template for Acme Ltd.” If the template has several workflows, the assistant asks which ones to start (in the card you choose). It then names the required fields of the first tasks and asks you for them. In Claude you can also say “Prepare the case first”: a card “New case” then shows the template, the values from your message, what opens next and which details are still missing — the case is created only when you click. “Create with open questions” also creates a task for you per missing detail; “Create” creates only the case.
  • Fill in fields: “On #227, in the task Review contract, enter 1 November as the contract start.” In Claude you can also say “Fill in the task Create offer on #227”: a card shows the task's fields with the values from your message already filled in. If there are several matches, for example two customers with similar names, you choose in the card. Nothing is saved until “Save · task stays open” (files at once); then you can complete it. Anything with an effect outside you confirm in Bricksta as always.
  • Complete or skip tasks: “Complete the task Request documents on #227.” If required fields are still missing, the assistant names them and asks whether to fill them in. Afterwards it says which tasks are open next.
  • Take over or assign tasks: “Take over the task Prepare quote” or “Assign it to Anna Smith” — by full name or e-mail address, as in Bricksta.
  • Set due dates: “Set the due date for Prepare quote to 24 October.” Customers do not see due dates.
  • Comment: “Write on #227: customer will call back on Monday.” Comments are internal; external parties in the portal or via guest links do not see them.
  • Turn a note into a case: “Muster GmbH wants two workshop days for €2,400, not three — create a quote for it.” The assistant creates the case in one go, with customer and positions — if a value does not fit, it creates nothing and asks. If something is missing, such as the contact person, it does not guess: it asks you or adds a task “Clarify the contact person” that is assigned to you.
  • Prepare approvals: “Which approvals are waiting for me?” and then “Approve the approval on #227.” You decide yourself — see “When an action reaches beyond Bricksta”. In Claude the waiting approvals appear as an “Approvals” card: the oldest in focus with the fields the template sets up for it, the others as rows; “Approve in Bricksta” opens it in the case.
  • Remind approvers: “Remind the approvers of the contract on #227.” This is the same as “Remind” on My work: at most once a day per task, only if you submitted the task or work on it, not for approvers outside your company. A second time on the same day sends nothing; the assistant tells you the time of the first reminder. In Claude, every waiting task on the “My tasks” card has a “Remind approvers” button, then “Reminded at 14:32”.
  • Remind the assignees of an overdue task: “Remind the assignees of the quote on #227.” This is the same as “Remind assignees” in the task sheet: only for overdue tasks that are not waiting for approval, and only if you own the case, have or had a task in it, or manage an assignee — just seeing the case is not enough. At most once a day per task; a second time sends nothing, and the assistant tells you the time of the first reminder. In Claude, the case card shows a “Remind assignees” button under an overdue task in “What it is waiting for”, then “Reminded at 14:32”.
  • Check what became of proposals: “What happened to my proposals?” — open, done, discarded or expired.

Every change shows in the history as “<name> (via Claude)” or “(via ChatGPT)”, comments included. On the Claude Team plan, the owner can additionally set “Complete a task” to “Needs approval”.

As an admin, ask “What was done via the AI chat?” to get the actions of the last 30 days (up to 90 days on request): who, what, when, in which case, through which assistant and whether directly in the chat, confirmed in Bricksta or in the card. In Claude this appears as a “Via the AI chat” card; each row opens the case. Only actions are counted — reading through the chat is not logged. Field values are not included. Members without admin rights get a refusal.

Assign several tasks at once

View documents in the chat

Attach a file or photo to a case

Put a file or photo into a file field

Which fields the assistant fills in

Text, number, yes/no, date
As in Bricksta.
Single and multiple choice
By the option's name; for multiple choice, several are possible.
Person
By full name or e-mail address of an active member.
Dataset field
By the entry as it appears in the dataset. This fills in the linked fields of the task; the assistant lists them with their values (“Also filled in: Description = …”). If the entry is not in the dataset or several match, it saves nothing and asks. If the field is restricted by another field, that one must be filled in first.
Fields hidden from the AI
The assistant does not read fields marked “Exclude from AI steps” and does not fill them in itself. Anything computed from them is hidden as well: a calculated field (formula or rollup) that uses a hidden field — across several steps too — and a group with one hidden column. If a dataset field fills them in, it only gives the count (“1 field that is hidden from the AI”) — you see the value in Bricksta.
Reference field
Shows the value of another field and is not an input. The assistant names that field and the task where you enter the value.
Table of cards (e.g. quote positions)
The assistant adds new cards, for example a position with service, price and project. Existing cards stay as they are. It checks every value like a single field (see Dataset field). Choosing a service from the dataset fills the linked values of the same card, such as the price.
File, signature, formatted text
File: in the “Fill in task” card. Otherwise only in Bricksta (link).

When an action reaches beyond Bricksta

Some actions reach people or systems outside your company: they open a task in the customer portal, send an e-mail or message, notify a connected system or change information external parties can see. The assistant does not carry out such actions itself. It creates a proposal, names the reason and gives you a link; the proposal also shows in the bell for 24 hours.

  1. 1
    In Claude a card “Confirmation needed” appears: it shows what goes outside — for example “Bricksta sends a completion email” with the task's name — and the email as it would be sent (recipient, subject, text). Details that must not reach the AI are hidden there. “Discard request · nothing is sent” discards the proposal right in the card; “Confirm in Bricksta” opens the page below.
  2. 2
    If your organisation has switched on “Confirm external actions in the chat” and you ticked the checkbox of the same name when connecting, the card shows “Confirm and carry out” instead. One click carries out the action right away, exactly as the card shows it. The button is valid for ten minutes and only appears in programs that show cards (e.g. Claude). If anything changed in the meantime, the card shows the new version and you check again. Reminder emails you still confirm in Bricksta.
  3. 3
    Four-eyes approvals have their own switch and checkbox: “Grant four-eyes approvals in chat”. With both on, the card shows “Approve” or “Reject”, who submitted it and the values of the approval. When rejecting you write the reason yourself; you take over a suggestion from the assistant only with your own click. No comment is stored when approving. The checkbox lasts 30 days; after that you reconnect.
  4. 4
    In chat this only works if you can see what you decide. The card leads to Bricksta and says why if no values are set up for the approval or values are hidden from the assistant, if an admin decides instead of the customer, for cancellation requests, if rejecting triggers further steps, and if the approval reaches outside your organisation while “Confirm external actions in the chat” is off. The history shows “in chat via Claude”.
  5. 5
    Open the link. The page “Confirm action” shows what the assistant wants to do, what reaches outside (with names from Bricksta), the completion email with recipient, subject and text and — for a field — the proposed value. Who a sending step writes to is often only known once it runs; the page says so.
  6. 6
    If a dataset field fills in further fields, they are listed under “Also changes these fields”.
  7. 7
    Choose “Confirm and carry out” or “Discard”. Nothing happens until you click. If what happens has changed since you opened it — another way outside, a different email —, Bricksta runs nothing and shows the new state. If you come from the card and something changed since, the page tells you.

The assistant never decides approvals itself. It prepares your decision; you decide with your own click in Bricksta or, if “Grant four-eyes approvals in chat” is on, in the card. The same rules as in Bricksta apply: the four-eyes principle, the approval right, a reason when rejecting. If a proposal looks unfamiliar, you disconnect with one click in Bricksta.

If the template sets up fields for an approval (in the template editor under “What the approval shows”), the assistant names them for every waiting approval and the approval shows them in that order — a field excluded from AI only on the page in Bricksta.

Step 1: Enable access for the organisation (admin)

Through Claude or ChatGPT, data from Bricksta goes to the respective AI provider: case data and — for people who may manage templates — also the content of templates, i.e. email texts, instructions for AI steps, fixed recipient addresses and webhook addresses. That is why access is switched off for every organisation at first, and only a person with admin rights can switch it on.

  1. 1
    Open Organisation › AI connections.
  2. 2
    Under “Access via AI assistants (Claude, ChatGPT)”, choose a level: “Off”, “Read only” or “Read and write”. The description below says what the level allows.
  3. 3
    Below it you now see all active connections: who is connected, with which assistant, since when and whether the connection may write. If you go back from “Read and write” to “Read only”, all connections lose the right to write; to write again, each person reconnects.
  4. 4
    Above the list, switch on “Confirm external actions in the chat” — off by default, only with “Read and write”. Members may then agree when connecting to confirm actions with external effect with one click in the chat card. Switching it off removes this permission from all connections; after switching it back on, each person agrees again when connecting. The list marks connections with the permission “confirms in chat”.
  5. 5
    Below it you switch on “Grant four-eyes approvals in chat” — also off by default, only with “Read and write”, independent of the switch above. It only applies to approvals with values set up in the template. The list shows for each connection until when the person agreed; the consent expires after 30 days.
  6. 6
    Below it you switch on “Attach files in the chat” — off by default, only with “Read and write”. Members may then agree when connecting to attach files and photos to cases from the chat card. Turning it off takes this permission from all connections; connections with permission show “attaches files” in the list.
  7. 7
    Below the switch, “View approvals without values” leads to a list of all approvals without selected values, each with a suggestion from the fields of the task and earlier tasks. “Apply” only saves to the template draft; it takes effect once you publish the template, and only for new cases. Fields from earlier tasks then also appear in the form of the approval task, read-only. The list doesn't offer fields excluded from AI or limited to certain roles, because the approval wouldn't work in the chat otherwise.

If “AI features of this organization” further up is set to “Off”, this access stays off too: switching AI off ends all connections, and “Read only” or “Read and write” can only be chosen again once the organization’s AI is on.

Step 2: Connect Bricksta in the assistant

  1. 1
    Claude (claude.ai or Claude Desktop): under Settings › Connectors, add a custom connector with the address https://bricksta.com/api/mcp.
  2. 2
    ChatGPT: in developer mode, create an app for a remote MCP server with the same address (available on the Plus, Pro, Business, Enterprise and Education plans).
  3. 3
    On the first call, the assistant opens the Bricksta sign-in. Sign in if you are not signed in yet.
  4. 4
    The connection requires two-factor sign-in. If you have not set it up yet, the consent page takes you to the settings; afterwards, start the connection again in the assistant.

The consent page shows which assistant wants access and where the data goes. A connection applies to exactly one organisation: the one you currently have open in Bricksta is preselected; you can choose another. Only organisations with access switched on are offered. For a second organisation, connect a second time. If the chosen organisation is set to “Read and write”, the tick “May write on your behalf” appears — unset by default. Without it, the assistant can only read.

If the organisation has switched on “Confirm external actions in the chat”, a separate checkbox for it appears below — off by default and only together with writing. It can only be set on this page, after two-factor sign-in. Approvals in chat have their own checkbox (30 days), and so does “Attach files” (no expiry).

Only Claude (claude.ai) and ChatGPT (chatgpt.com) can connect; the data goes only to the address shown. If you did not start the connection yourself, choose “Deny”.

Revoking a connection

A person with admin rights revokes any connection under Organisation › AI connections › “Revoke”. This takes effect immediately: the next question from the chat gets no more data. If an admin switches access off for the whole organisation, this also applies immediately to all connections. Anyone who loses their membership also loses the connection.

If you change your password or your two-factor sign-in, Bricksta ends all your connections to AI assistants. Connect again in the assistant afterwards.

If the assistant shows a message

Access not enabled
An admin has not (yet) switched on access for the organisation — see step 1.
Connection no longer exists
It was revoked or the membership ended. Connect Bricksta again in the assistant.
Data could not be loaded
A temporary error. Please ask again — it does not mean that there is nothing there.
Too many requests
Each connection may make 60 requests per minute and four at the same time, including 20 changes per minute. Wait a moment and ask again.
Writing not allowed
The organisation is set to “Read only”, or you did not tick the box when connecting. An admin changes the level; then reconnect and tick the box.
It is unclear whether it was saved
The answer did not come back in time. Check the case first before trying again — otherwise it may end up there twice.
Not carried out yet — please confirm in Bricksta
The action reaches beyond Bricksta. Open the link and confirm or discard it (see above).
Required fields missing
The task can only be completed once they are filled in. The assistant asks whether to fill them in.
Not in the dataset / several entries match
Nothing was saved. Name the entry as it appears in the dataset, or say which one is meant.
20 proposals already waiting
20 proposals from this connection are waiting for confirmation. Confirm or discard them in Bricksta first (bell).