Hey everyone, I’m Charles Benjamin, a CS undergrad from McGill and core contributor to the MOOS-IvP community.
I’ve been building MOOS-IvP Skills, a collection of agent skills for developing and validating software in MOOS-IvP, an open-source autonomy framework. For those unfamiliar, MOOS is a pub-sub middleware analogous to ROS, and IvP is the behavior-based decision layer that sits on top. MOOS-IvP is primarily used for autonomous marine robots.
Each skill covers a bounded workflow and hands off related work to other skills as needed. Right now, they cover MOOS app, IvP behavior, and mission creation, self-evaluating missions and multi-case harnesses, documentation lookup, log analysis, extension-repository setup, and georeferenced TIFF map generation.
For reference, here are the current skills and a graph of the handoffs between them
moos-ivp-mission-builder can use moos-ivp-docs to confirm behavior or app parameters in the initial construction process, then use moos-alog-analysis to inspect runtime evidence. Likewise, moos-ivp-eval-mission-builder explicitly starts with moos-ivp-mission-builder for ordinary mission layout before adding eval-specific headless launch, grading, results.txt, and completion functionality.
Version 1.4.12 is publicly available through the Codex plugin directory. Claude Code is also supported, and its public listing is pending review. Other compatible agents can install the skills directly from GitHub using the open skills CLI:
npx skills add https://github.com/cbenjamin23/moos-ivp-skills
I developed and revised the plugin throughout the summer through sandboxed evaluations and use in real projects like UAV/USV teaming at MIT and building a CI/CD pipeline for MOOS-IvP. I presented the work at MOOS-DAWG 2026, and the companion article goes deep into the design and project examples.
I also ran a 440-run benchmark comparing coding-agent work with and without the skills across eleven representative MOOS-IvP tasks. It looks at functional completion, engineering design, elapsed time, cost, etc.
I’d love feedback from MOOS-IvP users, robotics and autonomy developers, and people working on coding-agent tools. At this point I’m particularly interested in spotting potential new skills, awkward instructions, whether the skill boundaries and handoffs make sense, installation problems, ideas for evaluating the package, and whether a similar approach (to the plugin or subsequent benchmarking) would be useful in other specialized robotics frameworks like ROS 2, PX4, or ArduPilot.