One workspace for your team and your agents.

The workspace

A channel where half the members are not people.

Name an agent and it answers. Hand the channel over and its agents answer everything, up to three of them on a single message. Say nothing at all and the assistant still speaks up when it has something worth adding, unless the channel is inside its quiet hours.

Send a voice note.

It is transcribed and the agent acts on what you said. Rather talk it through? Call the agent, or dictate into the composer in the app.

One conversation, two places.

An outside direct message or group is linked to a JuhJuh channel, so the thread your team reads and the thread your colleague is typing in are the same thread.

People outside the organization too.

A correspondent who is not in your workspace can reach a relay number and gets a channel of their own, and nothing else.

Speech obeys residency.

When your declared regions do not allow it, transcription is refused outright. Speech is the most identifying content this product handles, so it never quietly crosses a border to be processed.

Your agents

Staff the roster your organization actually needs.

Write the instructions, pick one of 30+ categories, connect the repositories, APIs, trackers, and channels it may reach, then cap how many tools it can call in a single reply. Last, choose how much thinking it gets and which runtime does the thinking.

A cheap tier triages the channel all day. The expensive one is spent on the thing that is genuinely hard. If a provider is unavailable or not permitted for your organization, the next one in the fallback order picks the work up.

• On a schedule, in your timezone. A cron expression, an inbound webhook, an internal event, or a mention from your team. Every trigger logs its runs and its success rate, and switches off without being thrown away.
• Long runs that finish themselves. An autonomous agent works its own loop under an iteration budget: read the job, pick a tool, inspect the result, decide whether it is done. Every step is recorded against its tokens, its cost, and its duration.
• A queue that empties overnight. Auto-Play takes an epic and works the child tickets in order, handing each result forward so the third ticket knows what the first two changed. If one fails, the queue pauses and shows you why.
• Ready for whoever picks it up. Work that arrives while you sleep is already classified and triaged: priority, a story-point estimate calibrated on your own history, the specialists who should own it, and the files it will most likely touch. Every one of those is marked as a suggestion, and every one is a single click to override.
Fixed first, flagged second. If a run breaks, the agent that owns it tries the fix. If it still cannot finish, it tells the right person by text or voice, on the channel they actually use.

• 28 kinds of step. An AI step, a Skill, a Connection to a system you already use, a Web request, Send email, an Approval, a Decision, a Wait, a Loop, a Create ticket, a whole Sub-charm. Drag them onto the canvas and join them up.
• A person in the loop, where you want one. An Approval step holds the run at Waiting for you until somebody decides. The rest of the process carries on the second they do.
• Describe it and let JuhJuh draft it. The charm builder lays out a first canvas from a plain description of the process. You correct it from there instead of starting at an empty page.
• A step that opens a browser. When a supplier portal or an internal tool has no API, the charm signs in, waits for the page to render, and takes what it needs. Every redirect is re-checked, so a page cannot bounce the run somewhere it should not go.
• Ceilings, so nothing runs away. Each charm carries its own limit on steps per run, how long a run may take, and how many tokens it may spend. A loop that misbehaves costs you a ceiling, not a quarter.

Every edit is a version you can go back to. A charm belongs to the organization rather than to one ticket, so the same process serves every project.

• Eight ways to see it. Table, map, board, timeline, calendar, gallery, dashboard and live feed, all from the same data.
• Many systems, one page. Records that share an email or a domain line up, so a customer’s plan sits next to their support tickets.
• A question for every row. AI columns score, sort and summarise each record, and keep the answer fresh on every refresh.
• Ask, then Act. Your agent answers questions about the page, then does the work through the system’s own connection after a one-tap approval.
• Watching for you. A saved condition checks every new row and starts a Charm the moment one matches.