Agent Core

This project provides a Python SDK and agentcli command for the Agent Core runtime described in doc/.

The first implementation target is defined by:

• doc/21-codegen-contract.md
• doc/22-codegen-plan.md

The import package is agent_core, and the CLI command is agentcli.

Install

From the repository root:

python3 -m pip install -r requirements.txt

Create a default Ollama config:

mkdir -p ~/.soong-agent
cp src/agent_core/assets/templates/config_default.toml ~/.soong-agent/config.toml

Then run the TUI chat:

agentcli chat --path .

For a plain stdin/stdout loop, use:

agentcli chat --path . --plain

The default config uses local Ollama at http://127.0.0.1:11434 with model gemma4, so make sure Ollama is running and the model is available.

Agent Hub

Agent Hub is the local Electron/React desktop layer for agent_core.

Feature Demos

Create a conversation from a complex user request, send it to a configured team, watch agents coordinate through @mentions and an agent-managed Task DAG, then inspect the run result.

Configure a focused peer team for the conversation. Runtime and external agents are shown at the same Hub level, and not every agent has to join the team.

Load skills into the conversation context without turning them into teammates. Agents can then use the loaded skills while planning, building, and reviewing.

Message teammates directly with @mentions. User messages and agent-to-agent requests share the same activity stream instead of a separate comment/reply concept.

Run only the backend:

./agenthub --backend

Run the desktop app in development mode:

cd src/agent_hub/frontend
npm install
cd ../../..
./agenthub

If Electron cannot download its binary on your network, install with a mirror:

cd src/agent_hub/frontend
ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ npm install

The backend starts with PYTHONPATH=src python3 -m agent_hub.backend, creates ~/.soong-agent/config.toml only when it is missing, and uses the directory where you run ./agenthub as the project directory.