Context#

With only the skill installed, pi with a Codex model searched the web instead of rkb, and Claude called rkb through Bash without --toon. A skill loads only when the model picks it; a tool is in the tool list at every step. Claude Code adds model-visible tools only through MCP servers, and plugins cannot set permissions.

Decision#

Offer four tools (rkb_search, rkb_show, rkb_add, rkb_note), defined once in rkb and run by rkb tool <name>. Claude Code gets them from rkb mcp inside the rkb plugin, which rkb install claude writes as a local marketplace and installs with claude plugin; the rkb confirm ask rule and the tool allow rules stay in settings.json. pi and omp get them from the extension, a TypeScript package in extensions/rkb/ whose one source file is embedded in the binary. rkb mcp is hand-written blocking JSON-RPC.

Why#

Tools work whether or not the model loads the skill. One runner keeps the harnesses in step. The Rust MCP crates (rmcp, rust-mcp-sdk, brontes) all need Tokio, which rkb avoids, and the server needs only a few methods. Embedding the extension keeps it in step with the binary.

Rejected options#

• A single rkb tool that takes any command: harder for the model to use, and it would let an agent run rkb confirm around the harness gate.
• Loose files in ~/.claude (hooks in settings.json, command files, claude mcp add): no version and no single uninstall.
• Publishing the extension to npm: a second thing to version beside the binary.