AI Agents
Your sites, Base, and flows are where your business lives. AI Agents are the team that works it — they check your data, chase invoices, write reports, follow up with leads, and hand work to each other. They ask before anything big, show you exactly what they did, and can never overspend.
Build one in a sentence
Go to Automations → AI Agents → Build Specialist AI Agent and describe the job: “Every morning check my Leads base, fill in missing info, and flag anyone who asked about pricing.” AI drafts the whole agent — who it is, its instructions, the abilities it needs, and which of your sites and Bases it may touch. Review, adjust, build. You can also just tell the AI chat on that page to build it for you.
An agent only ever has the abilities and the sites/Bases you granted it. Everything else doesn't exist for it.
Asks first. Always.
Big actions — publishing, sending email campaigns, payments, deleting — pause the run as “Needs your approval” with Approve / Decline. That's the default; you can grant full autonomy per agent if you want. Every run shows a full report: each step, every tool it used, the exact instructions it was given, and what it cost.
A spend limit on every run
Runs on our built-in models use your plan's AI credits — the same monthly pool as AI chat — and every run shows its exact credit cost. Each agent has a per-run spend limit and a runs-per-day cap, and when your monthly credits are used up, agents stop instead of overspending. Prefer your own key? Agents can run on your OpenAI, Claude, Gemini, or other keys for zero credits.
They work without you clicking
- On a schedule — “every day at 9:00” in plain English.
- From your business — a new order, signup, or form submission can hand the agent a task automatically.
- From other apps — each agent can have a secret webhook URL.
- From each other — agents hand finished work to teammates you approved.
The Board — chained work
Chain tasks across agents: “Riley researches, THEN Piper writes it up.” Each task starts by itself the moment the ones it depends on finish, with the full work product handed over. Failed tasks show why and retry smarter; anything waiting on you shows in the Needs you column.
Skills & Memory
Skills are guides your agent follows — drop in any standard skill file (the same SKILL.md format as Claude skills; thousands exist publicly) or write your own steps. Agents can even save their own skills when they figure out a procedure that works. Memory is what they learn about your business — capped at 50 short facts (oldest rolls out automatically), and you can edit, switch off, or delete any fact.
For developers
Agents are fully available over MCP — Claude Code, Codex, or any MCP client can build, run, and manage them with the team_* tools. Backend functions can hand work to agents in one line: ctx.agents.enqueue('Agent name or id', task, data) — from account functions and from site functions (a site may only summon agents granted to it, and who may run the function is governed by its access rules).