Mandala and AI assistants.
2026 is the year assistants started doing things: sending, booking, buying, running commands. Some run on your machine, some in a vendor’s cloud. Mandala’s accomplice does those things too. The difference is what stands between the AI and the action.
Checked 12 October 2026
Four good products.
An open-source personal agent you run yourself and reach through the chat apps you already use: WhatsApp, Telegram, Slack, Discord, Signal, iMessage and many more. One-line install, apps on every major platform, memory kept as Markdown files, and a very large community.
An open-source agent that writes and improves its own skills as it works. One gateway for Telegram, Discord, Slack, WhatsApp, Signal and email, scheduled jobs, and a choice of places to run commands, from your machine to hardened containers and remote sandboxes.
Meta’s personal agent, launched in September 2026. Nothing to install: each person gets an isolated computer in Meta’s cloud, with connections to everyday services and a layer that asks before consequential actions. Available in the US at launch.
Always-on agents inside ChatGPT, announced in late September 2026. They keep pursuing a goal in the background, connect to a very large number of apps, and ask for approval before sensitive actions. Offered on some of ChatGPT’s paid plans at launch.
OpenClaw and Hermes Agent are described from their own repositories and documentation. Muse and Dots are described from launch coverage; we could not read the vendors’ own pages when we checked, so treat those two as reported, not confirmed.
Permission to use a tool is not a rule about your business.
Every serious assistant now has guards: approvals, allow-lists, sandboxes. They answer “may the AI run this command, or use this tool?”. That is necessary. It doesn’t know that a payment over $5,000 needs a second person, or that nothing ships until every crate is checked.
Mandala has those guards as well. Underneath them it has your rules, checked before any change to your records, for people and AI alike, and a signed record of what happened.
Plainly: this holds today for actions on your records. What goes out through a connected tool, such as an email, is held by approvals and that tool’s policy, like in other assistants. Bringing those under your rules too is on the roadmap.
This payment is over $5,000, so it needs a second approver. I’ve asked Leela. It goes out as soon as she approves.
Where they differ.
| OpenClaw · Hermes Agent | Muse · Dots | Mandala | |
|---|---|---|---|
| Where it runs | Your machine | The vendor’s cloud | Your machine: a Mac or your own server. No hosted edition |
| The model | Yours to choose | The vendor’s | Yours to choose, including a local one |
| What holds the AI | Command approvals, allow-lists and sandboxes that you configure | The vendor’s approval layer | Your business rules, checked before each action on your records; approvals for what goes out |
| Record of what it did | Session logs and transcripts | An activity trail kept by the vendor | Signed facts: who acted, for whom, under which rule |
| Memory | Markdown files you can read | Kept by the vendor, with controls to forget | Typed Markdown documents you can read and edit |
| What comes with it | An agent, its skills and its channels | An agent and its connections | An accomplice inside a workspace: mail and chats, documents, tasks, projects, records and dashboards |
| Who it serves | The person who runs it | The account holder | You, and the businesses you act for, by delegation |
| Channels and platforms | Many chat apps; Mac, Windows, Linux, phones | Phone apps, web, chat | Mail, Telegram, WhatsApp and voice; Mac and Linux servers today. iPhone, Windows and Android are coming |
| Licence | MIT, open source | Proprietary | Not published yet |
Said plainly.
OpenClaw and Hermes Agent talk through far more chat apps than Mandala does, and run on Windows and Android today.
Both have large open-source communities and libraries of shared skills. Mandala has a connector SDK and an MCP catalogue, and a much smaller world around them.
Muse and Dots need no machine and no install. Mandala asks you to run something, even if it is one program.
An agent with a shell can attempt almost anything. Mandala’s accomplice works through declared actions and tools. That is the point, and it is also a limit.
Choose honestly.
Choose an assistant if…
- You want an agent in the chat apps you already use, today, on any platform: OpenClaw or Hermes Agent.
- You want open source under a permissive licence and a large community to learn from.
- You want nothing to run yourself and are comfortable with your data in a vendor’s cloud: Muse or Dots.
- You want an agent for personal errands and code, not a system of record.
Choose Mandala if…
- The AI will touch money, orders, customers or anything you must answer for.
- You want rules that hold whoever asks, and a signed history of every action.
- You want the workspace as well as the AI: inbox, documents, plans, records, on your own machine.
- You act for a business, or several, and need “who acted, for whom” to be exact.
They are not exclusive. Mandala runs coding sessions with tools such as Claude Code and Codex on your machine, and can use any MCP server. How Prajna works · The AI answers to you
We checked every statement about other products on 12 October 2026, against their own sites, documentation and repositories where we could. Products change. If something here is wrong or out of date, write to [email protected] and we’ll correct it.