Agent Farm

Agent Farm MCP Server enables you to run tool-enabled agents that collaborate to perform complex tasks. It orchestrates multiple specialized roles, supports parallel work patterns, and can generate large documents or code by chunking work across workers. This guide walks you through practical usage, installation steps, and important configuration notes so you can start using the server effectively.

How to use

You operate Agent Farm through an MCP client that talks to the server to deploy swarms, run tool-enabled agents, and collect results. Start a swarm to assign tasks to colonies, monitor colony health, and trigger specialized swarms for file operations, code generation, or multi-perspective analysis. Use chunked workflows when you need large outputs like long documents or multi-function code files. The system automatically handles parallel work and assembles the final result.

How to install

Prerequisites: ensure you have Python installed and access to a shell. You will also use a virtual environment tool to isolate dependencies.

1. Prepare your project directory and environment.

2. Create and activate a virtual environment using the project’s tooling.

3. Install the server in editable mode to develop and run locally.

4. Run the server module to start listening for MCP client connections.

Configuration and runtime details

Agent Farm provides a standard runtime path to start the server locally from a virtual environment. The recommended runtime command is to execute the module that contains the server entry point within your virtual environment, for example: activate your virtual environment and run python -m agent_farm.server.

Available tools

spawn_colony

Create a new bug colony with configurable types (standard, fast, heavy, hybrid) to run a set of tasks.

list_colonies

List all active colonies with current health and task status.

colony_status

Provide detailed status for a specific colony including health, active tasks, and progress.

quick_colony

Run a quick health check on a colony to verify it is responsive.

dissolve_colony

Remove a colony and clean up its resources.

cleanup_idle

Remove colonies that have been idle beyond a configured threshold.

farm_stats

Return comprehensive statistics about the MCP environment and swarm activity.

deploy_swarm

Publish a set of tasks to a colony for execution.

quick_swarm

Spawn a one-shot swarm and deploy it immediately.

code_review_swarm

A 4-perspective review workflow for code, focusing on structure, patterns, quality, and performance.

code_gen_swarm

Generate code along with tests and documentation based on a spec.

file_swarm

Perform parallel file operations such as read, write, copy, or delete across multiple paths.

exec_swarm

Execute parallel shell commands across multiple agents.

api_swarm

Send parallel HTTP requests to target endpoints and aggregate results.

kmkb_swarm

Run multi-angle knowledge queries to gather diverse perspectives.

tool_swarm

Deploy bugs equipped with real system tools to perform tasks.

system_health_swarm

Quickly check the health of the host system and key services.

recon_swarm

Perform reconnaissance on directories or codebases to map structure and content.

deep_analysis_swarm

Deep disk/file analysis to extract actionable insights.

worker_task

Operate a single worker with access to full tooling for focused tasks.

heavy_write

Direct file write that bypasses large-language-model output limits to create large content.

synthesize

Standalone synthesis of any JSON results to produce a final consolidated view.