AGI-MCP

You deploy and run an MCP server that implements advanced Model Context Protocol capabilities, enabling persistent memory, subagent orchestration, lifecycle hooks, and a structured thinking process to execute complex tasks. This server is designed to be integrated with MCP clients to manage goals, observations, thoughts, commands, hypotheses, and assessments, while coordinating subagents and memory across sessions for robust, extensible AI workflows.

How to use

To leverage the MCP server, connect it to your MCP client using a native or containerized setup. You will configure the client to invoke the server for task analysis, ATLAS-driven task execution, and subagent collaboration. Your client can issue high-level goals, observe environmental state, and request actions through the server’s thinking and execution pipeline.

Typical usage patterns include: initiating a task with a goal, letting ATLAS break the work into subtasks, employing subagents to handle specialized domains, and querying the memory store for context across sessions. You will also configure hooks to customize the lifecycle events, such as pre-tool evaluation and post-tool results, to ensure safety and alignment with your objectives.

How to install

Prerequisites: you need a modern Node.js runtime and npm, and optionally Docker if you plan to run containerized instances.

Step 1: Clone the MCP server repository to your workspace.

Step 2: Install dependencies for the project.

Step 3: Build the project to produce runnable artifacts.

Step 4: Run tests to verify the build before deployment.

Optionally, you can run the server with Docker or use a local Node.js execution. Below are common commands demonstrated in practical setup flows.

# Clone the repository
git clone https://github.com/muah1987/AGI-MCP.git
cd AGI-MCP

# Install dependencies
npm install

# Build the project
npm run build

# Run tests
npm test

Configuration and starting tips

On first run, the server will prepare a memory documentation file, set up logs, and initialize the SQLite database. The server will also load built-in subagents and configure the hook system for lifecycle customization.

If you prefer containerized deployment, you can run the server via Docker using either the prebuilt image or a locally built image. The following examples illustrate typical start commands you would configure in your MCP client.

Note: exact command and path targets depend on your environment. Ensure you point to the correct distribution entry when running the native Node.js version.

Examples and quick references

The MCP client can be configured to connect to the server in either native or Docker form. Use the following configuration snippets as guidance for client integration.

{
  "mcpServers": {
    "agi-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/AGI-MCP/dist/index.js"]
    }
  }
}

{
  "mcpServers": {
    "agi-mcp": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "muah1987/agi-mcp:latest"]
    }
  }
}

Available tools

set_goal

Define system objectives with priorities to guide actions and planning.

observe

Record environmental observations to build an accurate state picture.

think

Reason about observations to form plans and determine next steps.

execute_command

Run and log commands as actions within the agent framework.

form_hypothesis

Create and test predictions based on observations and reasoning.

assess_performance

Evaluate outcomes and extract learnings to improve future runs.

process_goal_with_gotcha

Full cycle processing of goals through GOTCHA stages.

execute_atlas_task

Run the complete ATLAS 5-step task process.

get_atlas_history

Retrieve historical task execution traces for analysis.

get_active_goals

List currently active objectives within memory.

get_memory

Query memory by cognitive layer to retrieve stored data.

get_session_summary

Provide a summary of the current session and analytics.

execute_subagent

Delegate tasks to a specialized subagent.

resume_subagent

Continue work started by a subagent in a prior session.

list_subagents

Show available subagents and their statuses.