A just for fun personal LLM coding agent
Note: This isn't intended for widespread use. I developed it for fun as a first Gleam project and to experiment with the LLM APIs. It's not meant to be "production grade" and while I was cautious about destructive operations, I make no guarantees on its safety or operation.
A small CLI coding agent, providing:
- optional model selection (uses OpenRouter's API), defaulting to
anthropic/claude-sonnet-4.5 - basic tools (file search, file reading/grepping, file writing)
- basic safety mechanisms on these tools: checking for directory traversal, approval prompt on edits
- token counts on each message
Requires an OpenRouter API key to be present in the ROBOC_API_KEY environment
variable. Provide an alternate model selection with the --model flag.
Start the tool in the directory you want to work in.
$ roboc --help
Runs basic roboc agent
USAGE:
roboc [ ARGS ] [ --model=<STRING> ]
FLAGS:
--help Print help information
--model=<STRING> Which model to use on the provider. Defaults to
anthropic/claude-sonnet-4.5.- Sometimes the agent seems to get stuck in a "loop", repeating a previous message phrased in a slightly similar way
- Patch application is fairly unreliable right now, mostly in terms of the patches not applying at all due to malformed patch generation but other issues could arise too. Be careful!
- Other things potentially documented in TODO.txt
gleam run # Run the project
gleam test # Run the tests