Changelog

0.23.7 12 Oct 2026

Added

• The agent’s answers show text written as *italics* in italics, in notes too.

Changed

• The copy button now keeps only paragraphs, lists, bold and italics – no font, font size or colour. In your email program the text takes on its own font. Links become plain text.

0.23.6 12 Oct 2026

Added

• Every answer in the “AI” menu has a copy button at the top right. It copies only the content with paragraphs, lists and bold, without the grey background – so you can paste the text straight into an email.

0.23.5 11 Oct 2026

Changed

• The guide describes the new feed search (from Feed 2.3.0) and its permission “feed:read”.

0.23.4 11 Oct 2026

Changed

• The guide describes the new Feed tools: creating and publishing posts and polls (from Feed 2.2.0).

0.23.3 11 Oct 2026

Changed

• The chat in the “AI” menu now fits the window height: only the history scrolls, the title and input field stay in place. If you have scrolled up to reread something, a new reply no longer jumps the view to the bottom.
• The guide mentions the Calendar module’s new permission to create appointments (from Calendar 2.2.0).

0.23.2 11 Oct 2026

Changed

• The user guide in the module is up to date and shows three screenshots.

Fixed

• Under «AI approvals», proposals with before and after, for example for an FAQ article, are now shown line by line instead of on a single line.

0.23.1 11 Oct 2026

Fixed

• If your agent looked something up without further details (for example the list of mailboxes), the answer stopped with “http_400”. This is fixed.
• If the AI interface rejects a request, the reason is now shown instead of just a code.

0.23.0 11 Oct 2026

Changed

• McpConnect now follows the FreeScout module guide, like all other let's automate modules. Your settings are carried over automatically during the update.

Fixed

• If FreeScout runs in a subdirectory (for example under /helpdesk), all pages of the module now work. Until now they reported “not found”.

Security

• All outgoing connections (to the AI, to knowledge sources, to client metadata) now use the proxy and certificate verification configured in FreeScout.

0.22.3 11 Oct 2026

Changed

• On the agent page, «New service client» and «Edit» now open in a window above the page instead of always sitting in the page. The list of service clients stays easy to read.

0.22.2 11 Oct 2026

Added

• Tasks in the «AI» menu can be renamed.

Changed

• Your agent's picture in the «AI» menu is twice as large.

0.22.1 11 Oct 2026

Added

• In the «AI» menu your agent's profile picture now appears next to its name.

Changed

• The permissions on the service client and in the consent dialog are now grouped under headings: reading, notes, drafts and status, interventions, answering customers, knowledge base, and one block per module.

0.22.0 11 Oct 2026

Added

• Your other modules can now give your agent their own tools. Backup, Bexio and Subscriptions come first: your agent can check the state of your backups, look up contacts in Bexio and see subscriptions with their billings.
• Each module brings its own permission (for example backup:read). You grant it to your agent on the service client or to an AI client in the consent dialog.

Security

• Tools for admin data such as subscriptions or backups are only available when an administrator is behind the request – never when the agent is working on a customer message. Data from other modules counts as third-party content and is pseudonymised like tickets. Writing tools go through «Rights» and need approval by default.

0.21.0 11 Oct 2026

Added

• Recurring tasks in the «AI» menu: have a task repeat daily, on weekdays, weekly or monthly at a fixed time – for example “every Monday at 8 am: weekly report of open tickets”.
• The result appears in the task's thread and, if you like, also arrives by email. Administrators can also publish it in the feed (visible to everyone there).

Security

• Nobody is watching during a scheduled run: every change your agent wants to make waits for your approval. The recurring instruction is stored encrypted.

0.20.3 11 Oct 2026

Fixed

• On the «AI» page the main menu fell apart (the Themes module's custom order, groups and icons were missing), and answers did not load. Both are fixed.

0.20.2 11 Oct 2026

Added

• New permission tickets:close: your agent may close conversations when you ask it to in a note or in the «AI» menu – never on its own in response to a customer message. Under «Rights» you can set closing to approval or switch it off.

Fixed

• In the «AI» menu your agent's answers now appear by themselves, without reloading the page. Your message is sent without a reload.

0.20.1 11 Oct 2026

Added

• Tasks are deleted automatically 90 days after the last message. You set the period on the agent page; 0 means never.

Security

• Your tasks in the «AI» menu are now stored encrypted with the key of your installation. Anyone with only the database or a backup of it cannot read them. Existing tasks are encrypted during the update.
• The list of runs no longer shows any content for tasks, only tool, ticket and status.

0.20.0 11 Oct 2026

Added

• New menu item «AI»: give your agent tasks in a chat, without a ticket – for example “How many tickets about Bexio came in this week?” or “Which customers have been waiting for us for more than 5 days?”. Follow-up questions work in the same thread, and ticket numbers in the answers are linked.
• New tool for reports: count tickets, grouped by mailbox, status, assignee, day or week.
• On the agent page you decide who may give tasks (nobody, administrators or all users), how many messages per day are possible and how many changes per message are carried out directly.

Security

• In a task the agent only sees what both it and the person giving the task may see. Bulk actions always reach a person: beyond the set number, every change waits for approval. Everyone sees only their own tasks.

0.19.1 11 Oct 2026

Changed

• The list of recent runs now lives in the log, in the new «Agent runs» tab, with pagination. The agent page is much shorter as a result; a link remains there.

0.19.0 11 Oct 2026

Added

• The AI can correct, extend and create public FAQ articles in your FreeScout knowledge base (new permission kb:write, only when the Knowledge Base module is active).
• By default you approve every change and see the before and after. Under «Rights» you can switch to «direct». If the change was prompted by a customer email, your approval is always required.
• Every version is kept. Under «Knowledge» you see all changes made by the AI and revert them with one click; a newly created article is set to draft.

Security

• The AI does not touch articles with images, tables, videos or code. Links may only point to approved addresses, and customer data never ends up in a public article.

0.18.1 10 Oct 2026

Changed

• The «AI agent» dashboard tile has a new design: large numbers for today, this week and this month, a bar for spending against your daily budget, and the cost per answer and per chat below. It now fits the tile without scrolling.

0.18.0 10 Oct 2026

Added

• New «AI agent» tile for the Dashboard module: conversations handled today, this week and this month, average cost per answer and per chat, and spending today against your daily budget and this month. Visible to administrators only.

0.17.2 10 Oct 2026

Fixed

• The quiet period after an ended chat now survives updates and clearing the cache; before, the AI agent could answer the visitor again afterwards.

0.17.1 10 Oct 2026

Fixed

• A chat ended by the AI agent stays closed even if the visitor keeps writing: during the quiet period McpConnect sets it back to «Closed» right away. Their messages stay readable. Can be switched off under «Answering on its own → In its own words → Live chat».

0.17.0 10 Oct 2026

Added

• The AI agent can end a live chat when a visitor keeps asking private, unrelated or inappropriate questions after a friendly hint: it sends a closing message, closes the chat and does not answer this visitor for a while. Their messages still reach your team.

Security

• Ending works only in the chat, never for emails, and only after the agent has redirected once. Can be switched off; closing message and quiet period are adjustable under «Answering on its own → In its own words → Live chat».

0.16.3 10 Oct 2026

Changed

• In the chat the AI agent only writes an internal note when someone from the team needs to act, and always after its answer – so the note says what really happened.

Fixed

• A long chat no longer uses up the daily limit for emails: the live chat has its own limit (100 answers per day by default), adjustable under «Answering on its own → In its own words».
• What you release for sending in a note is no longer restricted by the AI agent's daily limit.

0.16.2 10 Oct 2026

Changed

• The evaluation «Is the threshold right?» separates email and live chat: for each channel you see from which confidence you sent drafts unchanged, and can lower that channel's threshold with one click.

0.16.1 10 Oct 2026

Added

• The AI agent can ask a short question when a request is unclear («What exactly would you like to know?»), in the chat right away. Such questions follow strict rules: short, no links, no amounts, no sensitive topics.

Fixed

• An old unsent draft of the AI agent no longer blocks answers: if the customer writes again, the agent replaces it – important above all in the chat.

0.16.0 10 Oct 2026

Added

• The AI agent also answers questions in FreeScout's live chat – with the same checks as for email, but without the visitor's email address and without waiting: the answer appears within a minute.
• The chat has its own threshold, delay and limit per chat, under «Answering on its own → In its own words».
• In the chat the AI agent writes briefly and conversationally, without greeting formula or signature.

Security

• The first answer in every chat starts with a notice that an AI is writing (EU AI Act, Art. 50). You can adjust the text.

0.15.1 10 Oct 2026

Fixed

• Security alerts from McpConnect that land in the support mailbox are no longer processed as tickets by the AI agent – saving cost and needless notes.

0.15.0 10 Oct 2026

Changed

• «Reply types» is no longer a menu item of its own but the tab «Templates» under «Answering on its own». Everything about when the AI agent may send on its own is now in one place: «General» with the delay for all sends and an overview, «In its own words» and «Templates».

0.14.1 10 Oct 2026

Fixed

• The number of actions in the list of runs and the notice about waiting knowledge suggestions are now translated.

0.14.0 10 Oct 2026

Added

• If you write «@Kiara send it» in a note, the AI agent sends its unchanged draft at your request – with the usual delay and «Stop sending».
• The AI agent learns from your clicks: the page «Answering on its own» shows how often you sent its drafts unchanged and suggests a suitable threshold.
• The delay can be 0: the answer then goes out within a minute.

Changed

• The decision «send or draft» is independent of the channel – the basis for the upcoming chat module with its own threshold.

Fixed

• Confident answers of the AI agent now really go out on their own: until now, the internal note it writes right after its answer stopped every automatic send.

0.13.2 10 Oct 2026

Added

• Form of address can be set under Agent → Runner: by default the AI agent addresses the team informally in notes and customers formally, unless they write informally themselves.

Fixed

• Times in the AI agent's notes match the time zone chosen in FreeScout under Settings → General.
• The number of open knowledge suggestions is shown right next to «Knowledge» in the menu.

0.13.1 10 Oct 2026

Added

• If you give the AI agent a fact in a note («all modules run on PHP 8.0 or newer»), it suggests a knowledge entry. You approve it under «Knowledge»; until then it has no effect.
• If the AI agent cannot take on a request by note, it answers you with a note and the reason.

Fixed

• Requests by note have their own daily limit per ticket and are no longer blocked by the limit for customer messages.
• Mentions from FreeScout's @ list (for example @kiara_(ki)) are recognised reliably, even after a rename.

Security

• Knowledge suggestions are only possible on behalf of a team member, never from a customer email, and contain no customer data.

0.13.0 10 Oct 2026

Added

• Address the AI agent in an internal note («Kiara, check the amounts again» or «@Kiara»): it carries out the request and answers with a note.
• It replaces its own reply draft as long as nobody has edited it yet; if you already adjusted it, your version stays.

Security

• Only team members with access to the ticket can address the agent this way; customer names in the note are replaced before they reach the AI provider when pseudonymisation is on.

0.12.0 10 Oct 2026

Added

• Entries of your own: facts in your own words, such as complete lists or prices, which the AI agent reads first and which take precedence over other sources – no Knowledge Base module needed.
• Settings per source: how amounts are read (for example in cents), whether customers get a link to it, and a note for the AI.
• Each draft shows what made the AI agent unsure, so you fix the cause under «Knowledge» instead of only the answer.
• Separate page «Answering on its own» for the switch, threshold, sensitive topics and figures; «Knowledge» now only holds sources and entries.

Changed

• The confidence of an answer now follows a fixed scale, which makes it easier to compare.
• Your existing list of sources is taken over on update.

Security

• A new or changed source while «Answering on its own» is on is reported to all administrators.

0.11.2 10 Oct 2026

Changed

• If the AI agent links a data address (JSON) instead of a page for people, the answer becomes a draft; when reading such data it is reminded that amounts are in cents.

Fixed

• Prices from data sources kept in cents, such as the shop catalogue, only count as verified when converted correctly; an answer stating «CHF 900» for a module costing CHF 9.00 becomes a draft.

0.11.1 9 Oct 2026

Changed

• If the runner is switched on but cannot work, the agent page and the sidebar show a red warning with the reason, and the log records skipped messages.

Fixed

• The AI agent also works when only admins may use AI connections under «Settings»; before, it silently stood still as soon as its service account had the role user.

0.11.0 9 Oct 2026

Added

• The AI agent answers questions in its own words, for example about your products – but only from sources you approve.
• New menu item «Knowledge»: approved addresses, optionally the public articles of the knowledge base, a source test and figures showing whether your threshold is right.
• Confidence threshold: below it, the answer becomes a draft that a person checks and sends.
• New tools «search_knowledge», «read_source» and «answer_customer», also for Claude Code.

Fixed

• Client ID metadata documents (CIMD) can be fetched again; outside the test suites the request always failed.

Security

• An answer only goes out on its own if every source it names was read, every link points to an approved address and every price appears in a source and is linked.
• Cancellations, refunds, invoices, contracts, complaints and legal questions are never answered on its own; you can add your own sensitive topics.
• Sources are only read over https, never from internal addresses, and redirects must stay on the list.
• Answering on its own is off by default and needs the new permission «answers:send»; switching it on notifies all administrators.

0.10.2 9 Oct 2026

Changed

• Reply types and sending have a menu item of their own, «Reply types»; the Agent page is shorter.
• «New reply type», «Edit» and «Delete» are buttons.

Fixed

• «Last used» shows the latest run for the AI agent's service client.
• A reply type without description can be edited again on PHP 8.5.

0.10.1 9 Oct 2026

Added

• Clients can be edited: name and redirect addresses, and a new secret for confidential clients.
• On the Agent page, «Edit» is now a button.

Fixed

• Confirmations such as «Stop agent?», «Empty list?» and «Approve?» appear again; they were missing on the Agent and Privacy pages and in the notice inside the conversation.

Security

• A new redirect address for a client notifies all administrators.

0.10.0 9 Oct 2026

Added

• The AI agent can send replies – only from reply types you release, and only by filling in their placeholders.
• Every reply is first a marked draft and only goes out after a delay; the conversation shows «Stop sending».
• New chapter «The AI agent» in the guide.

Security

• Sending is off by default; switching it on notifies all administrators.
• Replies go only to the customer, without CC, BCC or attachments, and never when someone already replied, a new message arrived or the email is suspicious.
• Whoever edits a scheduled draft takes it over: nothing is then sent automatically.

0.9.1 9 Oct 2026

Added

• Service clients of the AI agent can be edited afterwards: name, permissions, mailboxes, IP addresses and pseudonymization.
• A new secret can be created for each service client; the old one stops working immediately.

Security

• Every change to a service client ends its running sessions; new permissions only apply from the next one.
• When a service client gets more permissions, mailboxes or addresses, or pseudonymization is switched off, all administrators are notified.

0.9.0 9 Oct 2026

Added

• The AI agent can assign tickets to the right person in your team according to your house rules.
• New tool «list_team_members»: shows who in the team handles a mailbox, also for Claude Code.

Changed

• The list of runs shows for every action whether it was carried out, would be carried out or went for approval.

Security

• For suspicious emails, for example with hidden instructions or commands, the AI agent only writes an internal note and may suggest spam – no drafts, no assignment.
• The AI agent can never assign tickets to itself.

0.8.1 9 Oct 2026

Changed

• The list of runs shows the AI agent's notes and drafts as they would appear in the ticket.
• Reply drafts by the AI agent end without a name; FreeScout adds the signature of the person who sends them.
• The AI agent summarises each run in your team's language.

Fixed

• The Agent page no longer breaks on PHP 8.5 when a run ends without a summary.
• Costs in the log are rounded to four decimals.

0.8.0 5 Oct 2026

Added

• The AI agent is live: every new customer message is classified automatically, with an internal note and, if wanted, a draft reply.
• Trial mode by default: it writes nothing and only shows what would have happened.
• Your own Anthropic API key, house rules, limits per hour, day and ticket, plus an emergency stop.

0.7.0 5 Oct 2026

Added

• New 'Agent' menu item: service clients for an autonomous AI agent, bound to a service account, with browserless sign-in.
• Approvals for the agent go to the administrators.

Security

• An agent can only connect from registered IP addresses and can be stopped at once with an emergency switch. Its sign-in is valid for 15 minutes at a time.

0.6.0 5 Oct 2026

Added

• New 'Privacy' menu item: pseudonymisation and lists in one place.
• Additional first names, surnames, companies, streets, places, phone numbers, email addresses, custom terms and 'never replace' entries can be added or uploaded as CSV or JSON.

0.5.1 5 Oct 2026

Changed

• Pseudonymisation: FreeScout, the installation name, mailbox names and entries in a new 'never replace' list stay as they are.
• Known customer data is only replaced in its own conversation, not in other tickets of a result list.

0.5.0 5 Oct 2026

Added

• Pseudonymisation: names, addresses, phone numbers, email addresses, account numbers and similar data leave the server only as placeholders, for every AI provider.

0.4.2 4 Oct 2026

Changed

• Rejected client IDs are logged, and placeholders in the client form are marked as examples.

Fixed

• Safari no longer stalls after 'Allow access'.

0.4.1 4 Oct 2026

Added

• The self-check detects a web server rule that intercepts /.well-known/ and shows the missing line.

Fixed

• Right after installation, the self-check reports the scheduler as 'not run yet' instead of as an error.

0.4.0 4 Oct 2026

Added

• AI clients can register through a metadata document, only from approved domains and with strict checks.
• You can set which roles may have AI connections. When a role is withdrawn, its connections end immediately.
• Interface in French and Italian.
• The user guide now ships with the module.