Chat, Plan, and Execute use the same conversation but apply different tool policies. In the SoAI WebUI, open Chat from the sidebar, start or select a conversation, then select the current mode in the conversation header. You can also press Shift+Tab to cycle modes. Finish or stop the active turn before changing its mode or tool selection.
Choose modes and tools
| Mode | Use it for | Tool behavior |
|---|---|---|
| Chat | Questions, writing, analysis, and ordinary conversation. | Tools are optional and use the Chat selection. |
| Plan | Explore a task, clarify requirements, and write a plan without making changes. | Tools stay enabled, but destructive tools and tools without enough safety metadata are blocked. |
| Execute | Carry out a multi-step task with progress tracking. | Tools stay enabled and use the Execute selection. Selected tools can change files, run commands, control a browser, or affect an external service. |
Plan and Execute require a model with tool-calling support and at least one selected tool for that mode.
These modes also require the tool runtime to be enabled
(TOOLS.MCP.ENABLED).
Select tools by mode
In Chat mode, use Tools in the header to turn tool use on or off. Open Chat configuration > Tools to search tools, expand a server, and make separate selections for Chat mode, Plan mode, and Execute mode. Select Save. The picker includes built-in tools and available tools from connected MCP servers.
Tool selection requires
MCP usage
(MCP_USE). The header control is locked during a
turn. Indexed Knowledge also keeps its required retrieval tools enabled; disable Enable RAG under
Chat configuration > Knowledge before trying to turn those tools off.
Default tools (new chats) edits the staged tool lists for the open conversation and the defaults later chats inherit. Close that picker, review the main selections, and select Save to commit both.
Skills supply task instructions in ordinary chats and delegated work. The assistant can
load a selected skill with skill_read, or you can request an explicitly invocable skill with
$name. Explicit requests can supply instructions with tools disabled; any actions still follow the
conversation's mode, available tools, permissions, and approvals.
Approvals and checkpoints
Turn on Require tool approval under Chat configuration > Tools. SoAI pauses before a destructive tool, a tool marked as requiring approval, or a tool whose destructive metadata is missing or invalid. The panel names the tool and shows its proposed input. Select Approve or Deny.
Remember for this tool stores an account-level permission for later calls covered by the same rule. For most tools, that permission follows the tool name and does not narrow itself to the current arguments. Browser credential autofill is narrower: its remembered permission is limited to the destination site. If SoAI cannot determine that site, approve the current call without remembering it.
Plan mode continues to block destructive or insufficiently described tools even when an approval was remembered. To clear saved permissions, an Administrator opens the Settings button at the right of the header, then Settings > Reset > Tool approvals.
Respond to agent checkpoints
An agent turn can pause for one of three user responses:
- Tool approval asks whether a proposed call may run.
- Question requests one or more answers needed to continue.
- Secret request requests a credential and can offer to save it in the Secrets Vault with a label.
Answer the panel above the composer. Cancelling it or allowing it to expire produces a cancelled, timed-out, or failed interaction. The agent does not receive a fabricated successful tool result.
Monitor and control an agent turn
Plan and Execute can add a plan or todo panel above the timeline. After one exists, use Toggle plan bar or Toggle todo bar in the header to show or hide it. The progress events remain in the transcript.
In Plan mode, the agent can place its plan in the Plan widget and leave a short summary in chat. Select View plan to review it, then Execute plan when you want the agent to carry it out.
Tool rows can show state, input, result, duration, URL, code changes, media, or an error. Expand a row for detail and use View full output when it is available. A running shell action has its own Stop command. If an agent delegates work, subagent status, timing, and nested activity appear in the same assistant turn.
Stop, steer, or queue
During a turn, an empty composer can show Stop. When the current agent can accept guidance, text and ready attachments change the action to Steer. The agent takes a steer at its next step, such as after a tool call; while the model is writing its reply, the steer waits for the reply to finish. Select Stop or press Esc to end the reply and send the steer immediately. Otherwise SoAI can offer Queue for a later turn. A queued item can be cancelled until it starts. An approval, question, secret request, comparison turn, or existing queue item can restrict the next action; rely on the current button label.