These pillars are what turns a model into a living agent.
You govern each piece from the panel, and it's also within the agent's own reach.
Your agent's soul.
Unlike the memory of the main AI models, notes here are split into short and long term, and only what's actually needed gets loaded into context.
This saves tokens without compromising the quality of responses: the agent knows what it has stored long-term and consults it when needed.
The agent manages its own memory: it decides what to keep and for how long.
Read the docsYou Heads up: we no longer bill in 30 days. Since August, it's 15.
Kira · your agent Noted. I'll save it so it doesn't happen again.
the agent saves the correction to memory
Short term · what it knows now
Long term · what it remembers
One memory for each person
The knowledge it gathers from the person it's talking to is known in every private conversation with that person: the profile travels with the person, not the channel.
Conversations as context.
Every message in or out is logged and available to the agent. Search is multi-channel and lets you pull up entire conversations to resume context regardless of which channel you're on.
Threads are kept separate by person and by channel, so conversations don't bleed into each other.
Search is global and filterable by date: any past conversation can be found and picked back up exactly where it was left off.
Read the docsYou Do you remember what we talked about regarding March's quote?
doesn't remember, so it looks it up...
scope: your conversations and the groups you're in · all channels
Email Mar 14
«Send me March's quote with the volume discount.»
Telegram Apr 2
«It ended up being 12 pallets, not 10. Redo the calculation.»
Discord #purchasing Apr 14
«Management approves it. Invoice it within 15 days.»
finds it, loads it, and replies...
Kira · your agent Yes. You requested it by email on March 14th with a volume discount, in April you raised it to 12 pallets, and in #purchasing it was approved for 15 days on the 14th.
Self-managed, shared procedures.
A skill is a procedure that avoids redoing the reasoning behind repetitive tasks: billing criteria, a complaint-handling protocol, deployment steps.
They share a quirk with long-term memories: only what's actually needed gets loaded into context, and the agent consults it when it needs to. This saves tokens and keeps the model from getting overloaded.
They're assigned per agent and shared between several.
Each skill can have its own model and effort level, letting you tune cost procedure by procedure without touching the agent's configuration.
Read the docs
How a customer complaint is handled, step by step.
haiku-4-5 effort low
Issuing, numbering and who gets billed for each service.
haiku-4-5 effort low
Permission audits and key rotation.
opus-5 effort high
Steps and checks to ship a change.
agent's model
More than just chat
Telegram, Discord, WhatsApp, Slack, IMAP/SMTP email, Gmail, GitHub, webhooks. The agent operates on the channels your team already uses, with no extra tools to roll out.
Communication runs both ways: it handles what comes in and opens a conversation on its own initiative when something relevant happens. In groups, a required mention keeps it from butting in uninvited.
Agents can talk to each other, so one can hand off a task to another and get back to its own.
Read the docs
Share your credentials with your agents, securely.
Encrypted vault organized into cards by service: usernames, passwords, API tokens, SSH access and OTPs.
Assign each credential to agents and to your team. The value is never exposed in the conversation, and every access is logged in the audit trail: which credential, which agent, and when.
Read the docs
And every use is logged
Automate your work, or let it get done for you.
Inbox review, a closing report, monitoring a server, an anomaly alert. They run on schedule and leave a record of what they did.
Three modes depending on the job: invoke the agent to evaluate and decide, run a script and log its output, or combine both — the script gathers the data and the agent acts on it.
They support one-off or recurring runs, with an expiration date and their own time zone per task.
Read the docsA shared inbox where your team supervises live every conversation: which agent is handling whom and what it's replying.
Turns your agent into a public-facing operator without losing control: the team can supervise, step in and approve every sensitive action.
The agent can write drafts for the team to review, reply directly, or hand the conversation off to a human.
Luis I need the twelve pallets on Thursday, not on Monday. Can that work?
Yes. There's stock and the carrier has room on Thursday, so I've reserved it for you. The urgent shipping adds €180, and that charge isn't mine to approve: I'm passing it to Ana.
stock checked slot reserved route updated
GAZ hands the conversation to Ana Ana · support Hi Luis, it's Ana. Approved: it ships Thursday with urgent delivery included.
Words the wind won't carry off.
Send voice messages and make calls with your agents while you're on the move.
Each agent has its own assigned voice, and works the same way it does on any other channel.
Requires an integration with ElevenLabs or a compatible TTS/STT service.
Read the docs
Marta Hey, the order's up to twelve pallets. Can we still make Friday?
GAZ Yes. I'll tell production now and confirm as soon as they reply.
Marta Perfect, and note that the invoice is now fifteen days.
Nothing new on the horizon.
Connect and disconnect MCP tools on your agents: together with credentials, these are what extends what they can do.
Supports both stdio and HTTP transport. Each server's credentials are stored encrypted and never exposed: the agent uses the tool, never sees the secret.
Read the docs| Server | Via |  |  |  |  |
|---|---|---|---|---|---|
| github repository, issues and reviews | stdio | assigned to NOVA | assigned to SOKI | not assigned | not assigned |
| postgres ERP queries | stdio | assigned to NOVA | not assigned | not assigned | assigned to Kira |
| holded invoices and accounting | http | not assigned | not assigned | not assigned | assigned to Kira |
| notion team documentation | http | not assigned | not assigned | assigned to GAZ | assigned to Kira |
Fifteen days with everything unlocked, no card required. You just need a machine to run the agents on and the AI model you're going to use.