> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mythicsage.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Using Sage in chat

> Working with a Sage chat channel: choosing a mode, approving guarded actions, and stopping a running agent.

Every Sage session is created through Mythic's **New AI Chat** dialog. Channel creation and the
bot-token decision are covered in [Chat channel configuration](https://docs.mythicsage.ai/guides/chat-config.md); this page is
the day-to-day operating manual. Mythic renders
Sage output as markdown, and shows each tool execution as an updating, collapsible card.

## Pick a mode

Set `mode` when you create the channel, or switch it live with `/mode` (no argument reports the
current mode). Start in `conversation`, which is read-only; move to `supervised` when you want
Sage to propose actions you approve one at a time; use `auto` only for a prepared, scoped objective
in a lab. What each mode permits is defined in [Authority modes](https://docs.mythicsage.ai/understand/authority-modes.md).

## Reading the chips

Each Sage channel shows a row of status chips above the transcript. They report the channel's
live configuration and connection state without a model call:

| Chip | What it tells you |
| - | - |
| Provider, Model | Which model provider and model this channel uses. |
| Max Steps | The global cap on model steps for a run (default 200, `0` means unlimited). It halts a runaway loop. |
| Agent | The specialist currently executing your request, or `Idle`. |
| Policy | How Sage selects its next move. `hybrid` (the default) asks the model to choose only at real branches of an otherwise deterministic plan; `symbolic` always takes the first admissible action and cannot interpret Respond guidance; `llm` asks the model on every nonempty frontier. |
| Mode | The authority mode in force. |
| Autonomous | Whether deterministic multi-hop solving is forced for this session. |
| Mythic Tools, MCP Tools | How many tools are loaded, split between native Mythic tools and MCP-provided tools. |
| BloodHound, MCP Servers | Connection state: `connected` means the graph or external server is reachable. |
| Rounds | How many rounds the current session has run. |

You set provider, model, mode, policy, Max Steps, and credentials in the channel configuration
(the pencil icon on the channel header). Prefer storing the API key as a Mythic user secret
rather than in the channel config, which stores it in plaintext.

![A row of status chips: Agent Idle, Policy hybrid, Mode supervised, Autonomous off, Mythic Tools 26, BloodHound connected, MCP Servers 2, MCP Tools 32, Rounds 14](https://mintcdn.com/self-dbca2498/csYWoXextNaA1ISz/assets/chat-chips.png?fit=max\&auto=format\&n=csYWoXextNaA1ISz\&q=85\&s=51442771feea3390e3ce19e41f7a7ccf)

The chip row on a supervised Sage channel (Sage beta, captured July 2026).

## Approving guarded actions

In `supervised`, every guarded action (issuing a command, executing a capability, creating a
payload, adding a credential) pauses as a native Mythic approval card before it runs. Approval is
**default-deny**: only an explicit Accept lets it through. Read-only tools (`list_*`, `get_*`,
downloads) never prompt.

When the card shows a direct Mythic task proposal, it displays the deterministically prepared
command and arguments. If the command schema or selected registered file changes before issuance,
Sage stops and requires a fresh proposal.

If an accepted task has not returned a confirmed result by the wait deadline, the request ends
with the exact Mythic task ID and an incomplete-observation notice. The task's outcome is still
unknown; Sage does not silently re-issue it. A fresh request can inspect that task's output before
you decide whether to approve another action.

## Stopping a run

Two controls stop work, at different layers:

* **The composer stop button.** While Sage is streaming a response, the composer area shows an
  "AI response in progress" banner with a stop button to the right of where you type. Use it to
  cancel the in-flight request.
* **`/stop`.** Cooperatively stops the agent running on the channel. Use it when Sage is mid-run
  and you want the agent itself to halt cleanly.

The autonomous controller also enforces its own hard stop-losses (a cycle cap, wall clock, token
budget, and no-progress loop-breaker), independent of your Max Steps setting.

## What you can see

* `/state` shows the [engagement ledger](https://docs.mythicsage.ai/guides/engagement-ledger.md), the durable record of proven
  effects shared across chats in the same Mythic operation engagement.
* `/tools` lists the public Mythic tool inventory without a model call, distinguishing read-only
  tools from guarded actions.
* `/report` shows retained identity and membership observations and suggested next checks.

The full command list is in the [Slash commands reference](https://docs.mythicsage.ai/reference/slash-commands.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.