Monolith, a deterministic toolbox

If I give you a blank canvas and asked you to paint a landscape, your painting would be different from mine. Because language is squishy and broad creative prompts are meant to make room for the artist to express themselves.

Now image that I hand you a set of tools and explain each one: this brush is great for long strokes, for laying down the landscape; this one is great for clouds and detailing the branches of trees. Knowing what to use and when begins to nudge your process. Tools give us guard rails to follow, or push against, once you know how to hold them.

This process is something I've been thinking about more as I interact with llms. The wide gap between a starting prompt and the output. Give a coding agent a wide ranging tool like bash and they'll surprise you with the various ways they can approach the same problem.

You can constrain this behavior though by giving the agent dedicated tools, like search_files, read_file and write_file, each one with a dedicated use and strengths. This is the thinking that led me to create Monolith: a set of tools that agents can use via MCP or CLI to deterministically carry out requests.

Instead of one wide ranging tool, this tool box comes with a set where each one single purpose. This creates reproduceable steps and workflows, since the agent picks the exact tools it needs for the task. Each tool is chainable, following the UNIX philosophy, allowing you to create new pipelines for data processing.

This is an attempt at creating software for a computer to use. Looking at the where the current generation of coding agents fails and excels and giving them the tools they need to become more reliable.

-Mello

P.S.

I even made a fun animated ascii landing page for it: monolith.romellogoodman.com

P.P.S.

'Mysterious monolith' spotted for first time since 2020 in Nevada desert