Capability Router
Smarter tool selection for AI agents
About the project
AI agents that have access to many tools face a hidden problem. Every tool description loaded into context costs tokens, slows down reasoning, and increases the chance of the model picking the wrong one. When an agent has dozens of capabilities available, listing all of them upfront is wasteful. Capability Router was built to fix that.
The router works as a Codex plugin. Instead of loading every capability description into the active context, it scans the available tools, builds a compact local registry, and ranks them against the current task. Only the top candidates are returned. The rest stay out of context entirely.
A task comes in through a natural language instruction or the slash command. The router scores each capability against the task, applies soft masking rules to filter irrelevant options, and surfaces only what is likely to help. The agent then works with a focused set of tools rather than a noisy list of everything available.
The result is faster, cleaner agent execution. Less context noise means tighter reasoning and fewer wrong tool selections. It is the kind of infrastructure that does not change what an agent can do but changes how well it does it.