Tools overview

What tools are in Console, which kinds exist (MCP servers, custom HTTP tools, web search, page tools, sandbox) and how you bind them to an agent.

A tool is something an agent can call: a search, an API request, a browser action, a function in the page the user is looking at. The model decides when to call it, Console runs it, and the result goes back into the conversation.

The kinds of tool

Kind	What it is	Where you create it	Binding recorded as	Page
Custom tool	One upstream API imported from an OpenAPI or Postman document (or one remote MCP endpoint), flattened into a list of actions	Build → Agents → Tools	source: "tool"	Custom tools
MCP server	A full MCP server that Console hosts and proxies: from an OpenAPI spec, a remote MCP URL, an npx/uvx package, a Console capability (a Knowledge Base) or a composite of other servers	Build → MCP Servers	source: "mcp"	MCP servers
System tool	A built-in tool that needs one choice, not a definition: Browser Use (pick a browser) and Web Search (pick a Web Search instance)	Chosen in the Select Tools dialog	source: "system"	Web search, Browser
MCP Hub entry	A hub is a curated catalog of MCP servers for people and tools outside Console to discover. The servers inside it are ordinary MCP servers, so you bind those, not the hub	Build → MCP Servers → MCP Hubs (Enterprise)	the member server's own binding	MCP Hub
Page tool	A tool that runs in the host page when the chat is embedded in a portal: read the open order, fill a form	Declared in the portal, handled in your page	not a binding; declared per portal agent	Tools in your page

Agents also have a few capabilities that look like tools to the model but are switched on in their own Build sections rather than in Add Tools: asking the user a question and asking for approval (Approvals & questions), memory, the Knowledge Engine and the sandbox.

Bind tools to an agent

1. Open Build → Agents → your agent → Build.
2. In the Configuration card, choose Add Tools. The Select Tools dialog groups sources by badge: System, OpenAPI, MCP (custom tools of those types) and MCP (legacy) (servers from MCP Servers; the label is historical and this is the correct way to bind one).
3. Expand a source and tick the tools you want, or use Select all. For Browser Use and Web Search you pick the browser or the search instance instead.
4. Confirm writes the draft. Publish the agent for API and portal traffic to see the change.

Each binding stores the source kind, the source key, the chosen tool names and, for system tools, a small config (the browser or the search instance). The full walkthrough, with the form fields of each source, is in Build and publish an agent.

How a binding is resolved

A binding names its source by key, and keys are unique per project only. When a run starts, Console looks the key up in the agent's own project first and falls back only to a tenant-wide record of that key. It never falls back to another project's record, so an agent cannot end up running a same-named tool with another project's credentials. Saving the agent checks bindings the same way, so what validates at save is what resolves at run time.

A tool or server whose status is Disabled is skipped at run time even if an agent still has it bound. A single tool of an MCP server can be switched off too; see MCP servers.

Keep the tool menu small

Every tool left enabled is described to the model on every call, so an unused tool costs tokens each turn and gives the model one more way to choose wrongly. Bind only the tools an agent needs and switch off the rest.

Approvals

Tools that change something (a write, a shell command, a call to an external system) can wait for a person to allow them before they run. See Approvals & questions for the per-tool settings, who can answer and what happens when nobody does.

Call tools without an agent

Custom tools and MCP servers can also be called directly with an API token:

What	Endpoint
List tools / get one	GET /api/client/v1/tools, GET /api/client/v1/tools/{toolKey}
Run a tool action	POST /api/client/v1/tools/{toolKey}/actions/{actionKey}/execute
Run an MCP server tool	POST /api/client/v1/mcp/{serverKey}/execute

See the Tools API and the MCP API.