Configure a station — the target
Convoy decides who can do what, with which tools, in which order, when — and what happens next.
One station, every door, the same rules: our agent is just one client among others.
This version describes where we will be at the end of the roadmap. Its examples do not load in Core yet. For what works today, choose Current in the versions menu; for the path between the two, read From the current version to the target.
A station
A station is a deployment: the APIs of one or more tools, plus a config. A client — an organisation — owns one or more stations:
- a product client often has one: its product, made agentic;
- a teams client has one per process: "ticket → refund", "a new employee's arrival"…
Each station opens through MCP to any agent: Claude, ChatGPT, a company's agent. The agent
Convoy deploys — described at the end of the config, in the agent section — is a client
like any other, reachable on its portals: the widget and the PWA, Teams, Slack, WhatsApp. The
rules are the station's, for everyone.
| connected to | example | |
|---|---|---|
| Product | the APIs of the product being made agentic | a restaurant, a legal service |
| Process | the APIs of several tools | Zendesk + Stripe + Gmail |
A station can call another one: it lists it among its APIs, and the other station's rules apply on their own.
What the target adds
- Actions that read or write. A read runs right away and leaves a trace in the selection. A write is reserved, then executed at submit.
- Playbooks that are law. The order is enforced on every choice, and a playbook opens even if the agent does not ask for it.
- One identity per person: a room identified by data the person types, then authenticated by a provider (a code, an SSO, the tenant's login).
- Profiles: the same rules, adapted to who acts — a visitor, a customer, a company's agent, a manager, an automation.
- Everything is a policy: each restriction has a name, an effect, a mode, and is tracked in a registry. Smart policies, evaluated by a model, can only tighten.
- Declared events: incoming, internal or emitted. A fact is delivered; an action is decided.
- Who submits: a human by default, an agent where a rule says so, with its condition.
- Automations, approvals, journeys that last several days.
Who decides what
| actor | does | never does |
|---|---|---|
| the config | declares the tasks, the rules, the profiles, the thresholds | — |
| the decision point (Core) | says allowed, denied (reason) or a human must decide, and writes it to the log | delegate its answer |
| the planner | computes the open tasks and steers the model toward them | add a task missing from the config |
| the ranking model | ranks the candidates it is given; evaluates smart policies | authorise anything |
| the LLM | talks, fills in arguments, proposes calls | decide what is allowed or executed |
| a person | chooses, approves, submits | — |
A station works correctly with the ranking model switched off: it is only less smooth.