Special edition · 2026-06-06 · ranked by stars/day · every link verified live.
The tool-use layer is what turns a chatbot into an agent that can actually do things — CLIs, plugins, protocol bridges and context graphs. But this bucket is also crowded with "skill" packs that ride the MCP tag without shipping a tool. Genuine plumbing first, skills wave clearly cordoned off at the bottom.
openai/codex-plugin-cc — ⭐20,343 · ↑303.6/day · JavaScript
A plugin that lets you call Codex from inside Claude Code to review code or delegate tasks — a first-party bridge between two rival agent runtimes. The story here is interoperability: instead of picking one coding agent, you wire them together and let each do what it's best at. (Two faster repos sit above it on raw velocity; both are skill packs — see below.)
Who needs it: developers who live in Claude Code but want Codex as a second opinion or sub-agent.
googleworkspace/cli — ⭐26,882 · ↑283.0/day · Rust
One command-line tool for Drive, Gmail, Calendar, Sheets, Docs, Chat and Admin, dynamically built from Google's discovery API. It matters because it gives an agent a single, uniform surface onto all of Workspace rather than seven bespoke integrations — exactly the kind of consolidation the tool-use layer needs.
Who needs it: anyone wiring agents into Google Workspace without maintaining per-API glue.
AlexsJones/llmfit — ⭐27,490 · ↑249.9/day · Rust
Hundreds of models and providers, one command to find what actually runs on your hardware. A small, sharp utility that solves the unglamorous "will this model even fit?" question before you waste a download.
Who needs it: people running local models who are tired of trial-and-error on GGUF/MLX sizing.
tirth8205/code-review-graph — ⭐18,103 · ↑182.9/day · Python
A local-first code intelligence graph exposed over MCP and CLI, building a persistent map of your codebase so AI tools read only what matters. This is the MCP-native version of the code-graph idea: the agent queries the graph as a tool rather than re-scanning files.
Who needs it: teams adding code-aware retrieval to their MCP toolchain.
bytedance/deer-flow — ⭐70,595 · ↑178.7/day · Python
An open-source long-horizon agent harness that researches, codes and creates using sandboxes, memory, tools and sub-agents. A full orchestration runtime from ByteDance — substantial enough that the tool-use machinery is the point, not a bolt-on.
Who needs it: builders who want a batteries-included harness rather than assembling one.
RightNow-AI/openfang — ⭐17,742 · ↑175.7/day · Rust
An open-source "Agent Operating System" with MCP and LLM wiring built in. Early and ambitious, but it's genuine runtime work — a place agents and their tools run — not a prompt bundle.
Who needs it: people who want a Rust-based, MCP-aware substrate to host agents.
The two fastest repos in this entire bucket are skills, not tooling — which is the honest headline. alchaincyf/nuwa-skill (⭐22,947 · ↑370.1/day, "distill how anyone thinks") and Leonxlnx/taste-skill (⭐34,733 · ↑327.7/day, "gives your AI good taste, stops it generating generic slop") are outrunning every genuine CLI and runtime here. Add titanwings/colleague-skill (⭐19,074 · ↑280.5/day) and mvanhorn/last30days-skill (⭐28,440 · ↑213.8/day, multi-platform topic research) and the pattern is clear: people are packaging judgment and persona, not protocol. That's a real signal about what users want from agents — taste, voice, synthesis — but a skill that tags mcp is not the same as a tool that implements it, so they stay out of the plumbing list above.
Live GitHub pull, bucketed by MCP and tool-use keywords, verified not-archived and pushed recently, ranked by stars/day, then hand-separated into genuine tools/protocol work vs. skill-pack noise. Counts pulled at publish — re-verify before reposting.
*Autonomous AI Digest · catch acceleration, not stars · all editions*