Updated October 1, 2026

Agent mode and tools

What the AI can do in Agent mode, how to choose the connectors a chat can use, and how tool calls and approvals work.

In Agent mode, a chat node can act, not only answer. The AI can look things up in your connected apps, open other nodes on the canvas, and hand parts of a task to sub-agents, working through several steps before it replies. Every step shows up in the conversation, and changes to how the chat runs wait for you.

Agent is the default for new chat nodes. Switch to Chat (⌘2, or Ctrl+2 on Windows and Linux) when you want a focused reply with no tools, and back with ⌘1 (Ctrl+1). The Chat node page covers all the modes.

What the agent can do

Use your connectors. Search your inbox, read a shared document, check a project board: whatever you've added in Connectors. OpenAI models on your own OpenAI key can also search the web.
Read other nodes. A chat with no incoming connections (and Summarized context on) gets a short summary of the canvas and can open any node to read it in full. Once you connect nodes, it reads only what's connected. See Connections and context.
Delegate to sub-agents. It can split a task into parts, run each in a new connected chat node, and use their answers. See Sub-agents.
Propose a mode or model switch when your request needs one.

Agent mode doesn't build your canvas. Apart from the chat nodes its sub-agents run in, it never adds, connects, moves or edits nodes. That's Orchestrator mode, and the agent offers to switch when you ask for nodes.

Choose which connectors a chat can use

Every connector that's connected and switched on in Settings is on in every Agent chat. To turn one off for a single chat:

Click + (tooltip "Tools"), open Connectors and uncheck it, or
Type /, open Connectors and pick it. Each entry shows "On in this chat" or "Off in this chat".

The choice is saved with the chat node. Manage connectors opens Settings. In other modes the entry reads "Available in Agent mode".

If the AI reaches for an app you haven't connected, a Connect card appears with an Open Connectors button. Connect the app, then send your message again.

Note: Connector tools run as soon as the AI calls them, with every model. Nothing asks first, so turn off connectors that can send or change things in chats that don't need them.

Tool calls in the conversation

Each tool call appears in the answer as a row with the tool's name and a one-line summary. Click it to see what came back; long results end in Show more, and failed calls show the error. Two or more calls in a row fold into one line such as "Used 3 tools". Click it to see each call.

Approving tool calls

Two kinds of call wait for you: proposals to switch the chat's mode or model, and tools that Claude Code or Codex ask permission for. The row asks "Allow this tool to run?" and the AI pauses:

Approve runs it, and the AI carries on.
Reject skips it. The row notes that the tool didn't run, and the AI continues without it.
Always allow appears in two places. On a Claude Code or Codex request, it approves that tool for the rest of the current reply; your next message asks again. On a proposed switch to Orchestrator mode, it also approves future Orchestrator switches.

Mode and model switches

When a request fits another mode or model better, the agent proposes a switch. Ask for an image and it proposes Image mode; ask it to create or connect nodes and it proposes Orchestrator; ask for a different model and it proposes that model.

The proposal is a Switch Mode or Switch Model row with the target and a one-line reason. Nothing changes until you click Approve. The chat then switches and carries on with your request, so you don't need to resend it. Reject keeps the current setup.

Image mode needs a plan that includes image generation. Otherwise the switch is refused and an upgrade window opens. See Image generation.
The agent can move a chat into Chat or Image mode but not back out, since those modes have no tools. Switching back is up to you.

Auto-approve Orchestrator switches

Always allow on an Orchestrator proposal turns on Auto-switch into Orchestrator mode in Canvas settings. While it's on, Orchestrator proposals apply without asking, for every canvas. Orchestrator mode sends canvas context to Slashspace's cloud, so turn the setting off if you'd rather approve each switch.

Claude Code, Codex and Cursor

Subscription providers run their own agent loop with their own tools, such as editing files and running commands in the chat's agent workspace folder. In Agent mode Slashspace also gives them your connectors, and all three can start sub-agents through a built-in bridge. They don't propose mode or model switches.

Claude Code and Codex permission requests appear as approval rows. How often they ask depends on the permission mode or approval policy you set in Settings. A request left unanswered for 10 minutes is declined.

Whatever a tool returns becomes part of the conversation: it's sent to the chat's model provider (through Slashspace's hosted service for Slashspace models), saved in the chat's history, and readable by chats connected downstream. Keep that in mind with private email or files. See Privacy and your data.