Quick Start¶

This page shows the common setup path for a chat-style EAA agent:

configure an LLM
choose skills
register domain and built-in tools
optionally enable long-term memory
optionally launch the WebUI
run BaseTaskManager.run_conversation()

Run the script from a repository checkout after uv sync --all-extras. Set OPENAI_API_KEY in your shell first. The memory example uses the built-in Chroma-backed store, so it requires the memory_chroma extra.

import os
from pathlib import Path

from skimage import data

from eaa_core.api.llm_config import OpenAIConfig
from eaa_core.api.memory import MemoryManagerConfig
from eaa_core.gui.html import launch_html_webui_subprocess
from eaa_core.task_manager.base import BaseTaskManager
from eaa_core.tool.workspace import FileSystemTool
from eaa_imaging.tool.imaging.acquisition import SimulatedAcquireImage


PROJECT_ROOT = Path(__file__).resolve().parent
RUNTIME_URL = "http://127.0.0.1:8010"


def main() -> None:
    llm_config = OpenAIConfig(
        model="gpt-4o-mini",
        base_url="https://api.openai.com/v1",
        api_key=os.environ["OPENAI_API_KEY"],
    )
    memory_config = MemoryManagerConfig(
        enabled=True,
        persist_directory=str(PROJECT_ROOT / ".eaa_memory"),
        namespace="quick-start",
    )
    skill_dirs = [
        str(PROJECT_ROOT / "packages/eaa-core/src/eaa_core/skills"),
        str(PROJECT_ROOT / "packages/eaa-imaging/src/eaa_imaging/skills"),
    ]
    acquisition_tool = SimulatedAcquireImage(
        whole_image=data.camera(),
        add_axis_ticks=True,
    )
    workspace_tool = FileSystemTool(
        workspace_path=str(PROJECT_ROOT),
        read_whitelist_paths=skill_dirs,
    )

    task_manager = BaseTaskManager(
        llm_config=llm_config,
        memory_config=memory_config,
        tools=[acquisition_tool, workspace_tool],
        skill_dirs=skill_dirs,
        checkpoint_db_path=str(PROJECT_ROOT / "checkpoint.sqlite"),
        transcript_db_path=str(PROJECT_ROOT / "transcript.sqlite"),
        use_webui=True,
        webui_runtime_host="127.0.0.1",
        webui_runtime_port=8010,
    )
    task_manager.tool_manager.set_coding_tool_request_approval(True)

    task_manager.start_webui_runtime()
    webui_process = launch_html_webui_subprocess(
        RUNTIME_URL,
        host="127.0.0.1",
        port=8008,
        title="EAA Quick Start",
    )
    print("WebUI: http://127.0.0.1:8008")
    try:
        task_manager.run_conversation()
    finally:
        webui_process.terminate()
        task_manager.stop_webui_runtime()


if __name__ == "__main__":
    main()

For terminal-only chat, set use_webui=False and remove start_webui_runtime(), launch_html_webui_subprocess(), and stop_webui_runtime().

What the driver configures¶

OpenAIConfig is passed to BaseTaskManager.build_model().
MemoryManagerConfig(enabled=True, ...) enables retrieval and triggered saving in the chat graph.
skill_dirs controls the skill catalog. The same paths are passed to FileSystemTool(read_whitelist_paths=...) so the agent can inspect skill files without an approval prompt.
SimulatedAcquireImage is a domain tool registered through the task-manager tools argument.
FileSystemTool replaces the default filesystem tool handle with an explicitly configured workspace root.
use_webui=True creates the task-manager-owned WebUI runtime controller. start_webui_runtime() starts the agent-side API, and launch_html_webui_subprocess() starts the browser-facing server.

Slash commands¶

The base task-manager input parser supports these slash commands:

/exit: request exit from the active chat or task graph.
/return: return from chat to the caller or upper-level task when the chat graph is running.
/chat: switch from a task graph back into chat mode when a task graph boundary is waiting for user input.
/skill: list discovered skills.
/skill <name>: inject the selected skill's SKILL.md into context.
/skill <name> <message>: inject the selected skill and send <message> as the next user instruction in the same turn.
/setcodingtoolapproval true|false: toggle approval for the Python and Bash coding tools. The parser also accepts yes|no, on|off, and 1|0.
/setcodingtoolsandboxtype none|bubblewrap|container [visible_dir ...]: configure coding-tool sandboxing. Extra paths are used as visible directories for bubblewrap.

Unknown slash-prefixed input is treated as normal user text and sent to the model.