Greenhead - Agentic AI Framework in Go.

📢 Important Note on Project Status

Greenhead still has a lot of promise, because – AFAICT – it's the only framework that allows for config-only integration with your legacy command-line tools while also allowing for full customizability of the binary.

That said, Your Humble Author has other 🐟 to 🍳 at the moment, and further work on Greenhead is likely to be sporadic at best, unless someone forks it or hires me to customize the software to their needs.

In particular, considerations include:

• Finding a more flexible LLM API abstraction similar to Vercel AI
• Possibly moving to MCP (or maybe not; function calls are a good fit here)
• Rewriting the little web UI to be modern.
• Fix the access API, it's a mess.
• Surely other things besides!

Read on if you'd like to play with this or copy any of the ideas herein.

---

---

Greenhead is a framework for building and running AI Agents. It is a Go package, a command-line utility, a Web API, and a way of making custom command-line utilities and Web APIs.

There are four main ways to use Greenhead:

1. Run the ghd program to use any built-in or external tools.
2. Easily build your own program with any tools you care to define.
3. Use the Go packages to build anything you want.
4. Contribute tools for others to use.

Greenhead wants these use-cases to be as easy as possible while still being powerful enough to build Skynet.

Motivation

Many people have tasks they would like AI assistance with, but connecting the tools to Agents is often too difficult or insufficiently secure, and/or introduces permanent external dependencies.

Some of these tools can be defined as native functions; others already exist as binaries or scripts on the user's system.

Greenhead presents a uniform way of setting up both kinds of tools and presenting them to the AI. This, we hope, will make adoption of Agentic AI easier in situations where programming from scratch is not realistic, but full control of the environment is required.

Running the Command-Line Program

The standard CLI as shipped is named ghd and includes a standard --help option which is also available for every subcommand. In addition, detailed documentation is available via the doc subcommand:

ghd --help
ghd config --help
ghd doc
ghd doc config
ghd doc api

A good starting point is the prompt runner:

ghd agents run Howdy --log-file=tmp.log --agent=chatty --agent=pirate

Running the Web API

The Web API exposes persistent chat conversations over an HTTP interface. The server maintains chat context.

To run the HTTP API with two built-in agents available:

ghd api serve --agent=chatty --agent=pirate

Then direct your browser to localhost:3030, enter the temporary API Key printed by the above command, and click Submit Query to play with the primitive Web UI.

Don't forget to kill the server when you're done, John Connor!

Configuring External Tools

In order to expose your own programs as tools available to the LLM, you must configure them in the main (runner) config. Consider the UNIX host command, which should be readily available. Let's assume we want to use it to look up a host, optionally specifying the type and verbosity.

The usage string for this limited use of the command would be:

host [-v] [-t type] {name}

The config section would look like this:

[[external_tools]]
  name = "host"
  description = "Look up an internet host.."
  command = "/usr/bin/host" # or wherever it lives on your system!

  [[external_tools.args]]
    flag = "-t"
    key = "type"
    type = "string"
    description = "Query type: CNAME, NS, SOA, TXT, DNSKEY, AXFR, etc."
    optional = true

  [[external_tools.args]]
    flag = "-v"
    key = "verbose"
    type = "boolean"
    description = "Output verbose query results.."
    optional = true

  [[external_tools.args]]
    key = "hostname" # <-- Note: no flag specified here!
    type = "string"
    description = "Name of the host to look up, e.g. 'google.com'."
    optional = false

For more details, see the config [documentation][ghd] or use:

ghd doc config

Building Your Own

You can build your own version of ghd with very little code. Here is the "minimal" example:

package main

import (
    "github.com/biztos/greenhead"
    _ "github.com/biztos/greenhead/ghd/tools/tictactoe"
)

func main() {
    greenhead.CustomApp("minimal", "1.0.0", "SuperCorp Tic Tac Toe", "")
    greenhead.Run()
}

The pair run command could then be used to play a game of Tic Tac Toe between two OpenAI agents:

minimal run "Start game." --agent=tictactoe --agent=tictactoe -l tmp.log

The top-level greenhead package exposes functions for easy customization of your app. In addition to the name, you will usually want to customize the tools, and sometimes also the command-line options.

For more advanced customization, see the ghd/examples subdirectory.

Using the Packages

If your goal is to incorporate the agent-runner logic into your own project, you will eventually need to use more than the top-level greenhead package.

To copy the full run cycle, especially regarding config management, you will want to examine the ghd/runner subpackage. To manage agents and tools directly, the ghd/agent, ghd/tools, and ghd/registry packages should be consulted.

Subpackages in a nutshell, all under ghd/:

• agent - agent logic and API clients.
• api - the HTTP API; needs work
• assets - assets built from the src subdir with binsanity.
• cmd - the Cobra setup for the CLI; uses runner for command logic.
• registry - global registry of tools.
• rgxp - optional-regexp format for some config values. NB: may be factored out!
• runner - the command runner, including top-level config logic.
• tools - tools types and logic; subdirs contain the built-in tools.
• utils - misc utils; may be factored out at some point.
• version - canonical version numbers.

Contributing Tools

Tool contributions are very welcome! Pull requests will be evaluated first for safety and utility: it should be impossible for naïve users to cause damage, and the tool should be useful to a reasonable number of people. (Entertainment counts as "useful" -- games are encouraged!)

Please be sure that:

1. The tool name follows the namespacing conventions.
2. Descriptive text and documentation is (also) in English.
3. A working agent.config is included.
4. The README.md is easy for a layperson to understand.

For a canonical example, see the ghd/tools/tictactoe directory.

If you have an idea for a useful tool but are not a Go programmer, feel free to open a ticket describing your idea. Someone else may find it compelling and code it for you.

Namespacing

Tools may only have simple ASCII names like mytool_zebra_renoberator. The name should follow the subdirectory structure:

tools
└── mytool
    └── zebra
        └── renoberator
            ├── README.md
            ├── agent.toml
            ├── renoberator.go
            └── renoberator_test.go

A single subpackage may add multiple tools. For instance the renoberator package might define three tools:

• mytool_zebra_renoberator_obtain
• mytool_zebra_renoberator_stripe
• mytool_zebra_renoberator_destripe

These should be clearly described in the README file.

Especially for higher-level packages, avoid names that might conflict with future implementations at a lower level. The bias is for longer and more descriptive tool names.

External Binaries

Submitted tools may call external binaries, within reasonable safety bounds.

Such tools must check for the presence of the binaries before registering.

For some cases the easiest method is to wrap a standard ExternalTool in the normal tool packaging. This is legitimate, but the tool name must reflect any special usage or restrictions, to differentiate itself from other wrappings of the same binary.

For example a host command wrapped to take a -t option but nothing else might be named unix_host_lookup_by_type.

Acknowledgements

Thanks to Sam York for helping brainstorm this into liminal existance in Asia.

Greenhead is built with Cobra, Fiber, go-openai, and other great open-source software packages.

---