Chat Containers

A chat container receives an operator prompt, calls a model or tool system, and streams structured responses back to a Mythic AI chat channel. The Python mythic-container library provides the v4 chat base class and response helpers.

Minimal container

from mythic_container.ChatBase import (
    Chat,
    ChatConfigView,
    ChatModelDefinition,
    ChatModelMetadata,
    ChatRequest,
    ChatSecretView,
)

class MyChat(Chat):
    name = "my_chat"
    description = "Example provider-backed chat container"
    semver = "0.1.0"
    models = [
        ChatModelDefinition(
            Name="My Provider",
            Description="Streams responses from My Provider",
            Metadata=ChatModelMetadata(Provider="my-provider"),
        )
    ]

    async def chat(self, msg: ChatRequest) -> None:
        config = ChatConfigView.from_request(msg)
        secrets = ChatSecretView.from_request(msg)
        api_key = secrets.required_text("MY_PROVIDER_API_KEY")
        model = config.text("MY_PROVIDER_MODEL", "default-model")
        response_key = f"assistant:my-provider:{msg.RequestID}"

        await self.send_streaming(msg, response_key, metadata={"model": model})
        async for delta in stream_provider(api_key, model, msg.Prompt):
            await self.send_delta(msg, response_key, delta, metadata={"model": model})
        await self.send_complete(msg, response_key, complete_request=True)

One chat(msg) call handles one Mythic request. An approval or input continuation arrives as a new ChatRequest, not as a callback into the old coroutine.

Request shape

{
  "container_name": "my_chat",
  "operation_id": 7,
  "channel_id": 44,
  "apitokens_id": 91,
  "request_id": 1234,
  "model": "My Provider",
  "prompt": "Summarize active callbacks",
  "config": {"MY_PROVIDER_MODEL": "model-a"},
  "secrets": {"MY_PROVIDER_API_KEY": {"value": "..."}},
  "context": [],
  "slash_command": null,
  "input_response": null,
  "delegation_id": "",
  "delegation_name": ""
}

Never copy secrets or Bearer tokens into response content or metadata.

Streaming response shape

{
  "operation_id": 7,
  "request_id": 1234,
  "response_key": "assistant:my-provider:1234",
  "content": "partial text",
  "is_delta": true,
  "complete": false,
  "complete_request": false,
  "status": "streaming",
  "error": "",
  "metadata": {"provider": "my-provider", "model": "model-a"}
}

Reuse one response_key for updates to the same visible block. Use another key for tool cards, approval cards, or a separate final answer.

Human-in-the-loop and tools

Well-known metadata.special_type values let Mythic render native UI:

tool_use for tool progress and lazily fetched full output;
mcp_tool_confirmation for an approval before a write-capable MCP call;
input_requested for free-form or choice-based operator input;
subagent for delegated work grouped by delegation_id.

When a tool requires approval, finish the current request after sending the confirmation card. If approved, Mythic sends a new request containing the confirmed tool call. Rebuild typed configuration and run only that approved call.

Cancellation

The library tracks active request tasks and cancels them when Mythic sends a cancellation. Provider clients should propagate cancellation to their HTTP/streaming call and avoid sending a second terminal completion after cancellation. For a complete reference implementation, use the basic_chat example and the CHAT_CONTAINERS.md guide in the MythicContainerPyPi repository matching your installed library version.

​Minimal container

​Request shape

​Streaming response shape

​Human-in-the-loop and tools

​Cancellation

Minimal container

Request shape

Streaming response shape

Human-in-the-loop and tools

Cancellation