108 agents in 12 categories
Each one carries a method in its instructions: the steps it follows, the questions it asks when the brief is thin, what it hands over, and what it will not do. The sources it leans on are named, with their licence.
What happens between a request and a result
A team of agents takes over one repeatable process. Every run drafts a plan before it spends anything, and every correction you write becomes a rule the next run follows.
The loop
The same five steps every time, whether the request came from a form, an inbox or your own backend.
A request arrives
From the app, from the project's own email address, or from a signed HTTP POST to its webhook. Each channel can be revoked on its own.
The team drafts a plan
Steps in order, and which agent owns each one. Nothing has run yet and nothing has been spent.
You approve, edit or reject
The run waits for a person. Teams doing routine work can approve confident plans by themselves and still stop below the confidence threshold you set.
The team executes
Agents work the steps with the tools that team is allowed to use. Files they produce stay with the run, and you get an email with a link when it finishes.
You say what was wrong
Your comment is turned into one imperative sentence and handed to later runs — for this project, for the whole team, or for that agent everywhere it works.
What is in the box
Agents you write yourself
Hire one from the catalogue and rewrite its instructions, or start from an empty agent. Nothing you hire stays tied to its original.
A knowledge base per agent
Files, pasted text or a URL. The content is split, indexed and searched during a run, so the agent works from your material.
Tools over MCP
A team can use only the MCP servers it has been given. Credentials are held against your organisation, encrypted at rest, and connected once rather than per agent.
Runs you can read
The plan, the steps, the tool calls and the errors are kept in order, so you can see what the team did and where it stopped.
An API, and an MCP port
Start runs and read results from your own systems over a documented HTTP API — or let another agent drive SmartStaff over MCP.
What these agents are not
An agent is a set of instructions with a name attached. It is not a person, and we do not draw it as one: agents get abstract marks rather than portraits, because a convincing face would suggest somebody is standing behind the work. Nobody is. It is the same reason every run stops for a human by default.
Put a team on the work that repeats
Define the process once, and it runs every time the work arrives — with a plan you approve before anything happens.
Free plan available. No card needed to start.