Changelog

Mihu Loop now fits in your pocket. Your inbox, calls and tickets go wherever you go, so a customer never waits because you stepped away from your desk.

Sign in with your Mihu account and pick up right where you left off on the web or desktop.

• → iPhone and Android: download it from the App Store or Google Play.
• → One account everywhere: the same conversations and team on your phone, the web and the desktop app.

Get your custom boards. The first page you see after signing in is now yours to arrange: pick what matters to you, drop it in, and leave out the rest.

Not sure where to start? Pick a ready-made board for your business and your role, and change it from there. You can always go back to the default.

• → Build it like Lego: calls, chats, appointments, tickets, your AI agents, your team, your own reports. Pick the pieces you want, put them wherever you like, and build the board that works for you.
• → A starting point that fits you: ready-made boards for your industry and your role, so a sales manager and a support lead each start from something useful.
• → As many boards as you like: one for the morning check, one for the end of the month. Name them and switch in one click.
• → Do things from the board: make a call, send a WhatsApp message, connect a channel or create a new AI agent without leaving the page.
• → Your menu, your way: drag the menu into the order you like and hide what you never use. Only your menu changes, and every board you save shows up there as a shortcut.
• → Helpful when it is empty: instead of a zero, each box tells you what to do next, like getting a phone number or adding your first document.

Type what you want to see. The AI reads your workspace data model, builds each widget, checks its own numbers and lays them out on a drag-and-drop canvas. A report assistant takes follow-up instructions.

Eighty ready-made sector templates open instantly with your data, heavy results become Excel files by email, and every widget runs read-only against an allow-list of tables.

• → Verified widgets: each widget is built, run and then checked by a second pass that explains what it shows; problems are repaired before the widget is added.
• → Report assistant: “split by channel”, “last 7 days instead of 30”, “add average call duration”, “make the first chart a pie” — applied to the report you are looking at.
• → Canvas and filters: twelve-column layout you own, ten chart types, named parameters that become a filter bar with saved presets.
• → Templates: eighty reports across eight sectors built from a hundred-plus prepared widgets, no AI wait.
• → Excel by email: a pre-flight estimate catches heavy widgets and offers a background export with a signed-in download link; an Exports tab lists them all.
• → Read-only by construction: allowed tables only, hidden personal columns, single read queries inside a read-only transaction with a time limit.

Web Chat just got smarter. Your website agent now has a job to do there. Give it a goal, let it make offers within the buffer you allow, and let it learn from your visitors.

It recognises your customers from the same unified contacts as your calls, WhatsApp and email, and you build its first screen by drag and drop.

• → Goals: sell, capture leads, book appointments, recover carts, support, collect feedback, route to your team. Or let it read your site and suggest goals.
• → Offers within a buffer: give it a coupon and a maximum discount, and choose whether it decides when to use it or only offers it when asked.
• → Aggressive, Balanced, Strict: one choice sets how proactive it is, how it uses discounts and when it hands over to a person.
• → Learns from visitors: knows which page they are on, can personalize by visitor context, and every chat feeds Auto-Improvement.
• → Unified contacts: forms and signed-in customers are matched to the same contact as their calls and WhatsApp, so the agent knows who it is talking to.
• → Drag-and-drop first screen: contact form with your own fields, appointment picker, discount code, WhatsApp cards, quick options, consent notice, online advisors.
• → One script tag: design it on your own site, publish, and roll back any version in one click.

The agent reads its own conversations on every channel, collects what it could not answer or do, and turns repeated gaps into a short to-do list for you: knowledge, a data connection, a permission, a behaviour rule, or a fix.

Every request carries the conversations that caused it. You answer once, or you let the agent improve itself with Prepare answers or Apply automatically.

• → Five kinds: Knowledge, Data connection (intent), Permission, Behaviour, and Fix — the last built from failures the platform recorded, with no language model involved.
• → Evidence first: a request needs at least three conversations behind it, personal details are masked before analysis, and there are no suggestions without a customer having asked.
• → What an answer becomes: knowledge text, a document, a table row, an intent with a connection contract, a setting with its limits, or a guideline. Your prompt is never edited.
• → Automation: Off, Prepare answers (draft, you approve in one click) or Apply automatically (knowledge, read-only lookups and guidelines applied and versioned). Permissions always wait for a person.
• → Did it help: each answer reports how many conversations used it, and the page counts what the agent learned this month.

There are many ways to change your agent now: the dashboard, the API, the CLI, Claude or ChatGPT over MCP, Mihu Assistant, Simulator Studio, the agent’s own self-improvement. We built it that way on purpose. But we have to be honest: it was not clear what changed, where, when or by whom.

So we fixed it. Every real change is now saved as a version with its source, you can open any version and see exactly what changed, and you can go back to any of them in one click.

• → The whole agent: details, intents, procedures, guidelines, training, knowledge, voice and messaging, rules and routing, integrations.
• → What, where, when, who: every version says which parts changed, where the change came from and who made it.
• → See the difference: old on the left, new on the right, only the parts that changed.
• → Going back never deletes: a restore adds an entry to the history, so you can always move forward again.
• → Name the versions you trust: Approved by legal, Working well. Also available through the API and the CLI.
• → Connections are referenced, not copied: integration accounts and their credentials are not part of a version; secrets on the agent itself stay encrypted and are never shown.
• → Nothing to set up: on for every agent, on every plan.

Some work belongs in a terminal, a script or a CI job. The Mihu CLI puts every part of your workspace behind a single command: AI agents, phone numbers, calls, conversations, campaigns and more.

It also connects your AI coding tools to Mihu in one line, so Claude Code, Cursor or VS Code can manage your workspace in plain language.

• → One-line install: macOS, Linux and Windows, a single program with nothing else to install.
• → Every API endpoint is a command: the commands follow the API, and create or update commands ask for each value when you leave them empty.
• → Several workspaces: log in once per workspace and switch with one flag; use environment variables on build servers.
• → AI tools in one command: mihu mcp setup connects Claude Code, Claude Desktop, Cursor, VS Code and Windsurf to your workspace.
• → Made for scripts: readable tables for you, JSON for your scripts, and every page of a list at once.
• → Always current: mihu update, shell completions, and every release published on GitHub.

Mihu Loop is no longer just a tab in your browser. Download it and it runs directly on your computer — the same inbox, calls, and tickets, in its own window.

• → Native desktop app: for macOS, Windows and Linux — the download page detects your system and gets you the right build automatically.
• → Live in seconds: install, sign in, and your queues, conversations and tickets are exactly where you left them.
• → Always current: Mihu Loop checks for new versions on its own, so you're never behind.

The moment a conversation ends, Observer agent decides whether it needs a ticket — a complaint, a sales opportunity, or a loose end, matched against what you've told it to look for in Settings. If it does, it creates a ticket for your team or your queue to handle.

• → Auto-classified: tell Observer agent in plain language what should count as a complaint or an opportunity in Settings, and every ticket gets tagged and scored against that.
• → Handed to your team: each kind of ticket goes to the team member or queue you chose for it, and they're notified right away. Nobody chosen? It goes to your default queue.
• → Linked to the conversation: every ticket opens straight back to the original call or thread, no searching required.
• → List or Kanban: Open, In Progress, Waiting on Customer, Resolved, Closed — filter by priority, tag, source, or who's holding it.

It retrieves, it writes, it sends — and all three are on the record. Autopilot, Copilot and Human First can be set on a queue, on a channel, or on one conversation.

Copilot now reaches live calls too: the AI agent stays on the line, hears both sides, and prompts your team member with the next line instead of speaking itself.

• → Autopilot — the AI agent sends: it retrieves, writes and sends the reply itself, with every step on the record.
• → Copilot — it drafts, your team sends: the AI agent writes the whole reply, then stops at your team member's screen as an AI draft.
• → Say Next on live calls: the AI agent listens on the line and feeds your team member the next line to say, without ever speaking itself.
• → Take over, hand back: control can move between the AI agent and a team member in both directions, mid-conversation.

Sixteen thousand rows, and a score on every one of them. That score is the whole product.

Everything a floor needs, on one screen that never navigates away — one inbox, seven channels, and three columns: what is waiting, what you are doing, and what the AI agent knows.

When the AI agent stops, it does not go quiet — it raises an alarm, and every escalation state looks different on purpose so nothing gets missed.

• → One inbox, seven channels: every channel lands on the same desk, so nobody switches tabs to keep up.
• → Three modes, one switch: Autopilot, Copilot and Human First, set per queue, per channel, or per conversation from the header.
• → Voice, live: the AI agent joins the call itself and works alongside the team member on the line.
• → Escalations you cannot miss: a distinct state for every reason an AI agent hands a conversation back.

Set the mode on a queue, on a channel, or on one conversation. The customer never sees the switch.

Who gets into which part of Mihu, and how far — plus a full log of every screen opened and every request made, allowed or denied.

Team Management is also where queues route incoming calls and messages: one email, one role, and a new teammate is in.

• → Roles and scopes: bundle the exact permissions a job needs into one role, instead of ticking boxes for every new arrival.
• → Access logs: every request logged with the time, the member, the path they opened, and whether it went through.
• → Queues with SLA, flow and hours: set how fast, what happens off-plan, and when the lights are on, per queue.
• → Smart routing: rules of your own decide exactly where a call or message goes the moment it arrives.

Every queue in one line: its channels, its agents, its target and its rules.

The address people type, the lock on the door, who gets a key, the name on the wall — four settings, one sitting.

Every workspace still ships with a mihu.ai address that keeps working, but the workspace can now look and feel entirely like your own.

• → Your own domain: point an address you own, like app.yourcompany.com, at your workspace.
• → Workspace-wide 2FA: one switch turns on two-factor authentication for everyone in the workspace at once.
• → Hand out keys, not copies: publish a registration link and approve every access request yourself.
• → Put your name on it: swap the logo so your team opens a workspace that looks like yours.

Call, message, and track every lead without leaving Zoho CRM. Link your Mihu AI account and voice/WhatsApp agent once, and it's available on every record.

A lead comes in from a website form, a rep calls straight from the record, and the outcome is already on the timeline before they've opened a new tab.

• → Actions on the lead record: call, send a WhatsApp template, or add to a campaign in one click, without leaving the CRM.
• → Automatic logging: every call and message is written back as a note automatically, with outcome, content and timestamp.
• → Minutes to set up: connect your workspace subdomain and API token, choose your agents — no code required.