Skip to main content
The Termux Multi-Agent system provisions and orchestrates autonomous agents inside Termux on Android. It reads tasks from a central task registry, dispatches them through environment variables, and provides a live dashboard for monitoring agent state. The system also includes a CedarScript MCP server for external tool integration.

​
Core components

The multi-agent system lives in the termux-multi-agent/ directory and contains the following modules:
  • run.py — Main orchestration loop that initializes the database and starts the agent scheduler
  • provision_agent.py — Provisions a new agent with a unique identity, workspace, and environment
  • dashboard.py — Live terminal dashboard showing active agents, task queue, and completion status
  • patch_files.py — Applies file patches generated by agents during task execution
  • cedar-mcp-server.js — CedarScript MCP server that exposes agent capabilities to external clients

​
Task management

Agents receive tasks through environment variables rather than interactive prompts. The orchestrator reads the master task list from ~/workspace/llm_map/master_tasks.json and binds each task to an agent via the following variables:

​
Running an autonomous agent

Follow these steps to provision an agent and run it against a task from the master registry.
1

Initialize the database

From the termux-multi-agent/ directory, run the initialization script. This creates the SQLite database that tracks agents, tasks, and outcomes.
Run this once. The database persists across restarts and holds agent state.
2

Review the task registry

Inspect ~/workspace/llm_map/master_tasks.json to find the task you want to dispatch. Each entry contains a task_id, goal, target, and optional constraints.
3

Provision an agent

Use provision_agent.py to create a new agent. The script assigns a workspace under termux-multi-agent/workspaces/ and writes the environment file.
4

Set environment variables

Export the task variables so the agent can read them at startup. You can also write them to the agent’s .env file.
5

Start the agent

Launch the agent run loop. The agent reads its environment, loads the target file, performs the refactor, and writes results to TASK_WORKSPACE.
6

Monitor with the dashboard

In a separate Termux session, open the live dashboard to watch progress.

​
CedarScript MCP server

The CedarScript MCP server (cedar-mcp-server.js) exposes agent operations to external clients through the Model Context Protocol. Start it when you need to integrate the multi-agent system with an IDE or another orchestrator.
The MCP server requires Node.js to be installed in Termux via pkg install nodejs.

​
Next steps

Autonomous Tasks Guide

Learn advanced patterns for dispatching and chaining agent tasks.

First Session Guide

Walk through a complete ArchWiz session with agent integration.