🛠️ Claude Reads the Contract. You Keep the Money.

It's Friday, which means two things: the kit drops, and I found something worth looking at first.

I've been watching the Claude Code skill ecosystem accelerate this week. The anthropics/skills repo crossed 146,000 stars and operators are building on top of it fast. The community repos surfacing alongside it, things like ccundo for granular undo and SkillForge for intelligent skill routing, tell you where the real tool-building energy is right now. It's not in the big model releases. It's in the layer that makes Claude Code stickier and more precise for operators who already live in it.

Also worth naming: Vercel updated its legal terms this week specifically to address agentic workflows. Developers granting AI tools direct infrastructure access is now a fact Vercel is writing into contracts. Which is a good segue.

The clause that lets your client keep your work without paying you is not hidden. It's just written to be skipped. Most freelancers and solo operators sign contracts they don't fully understand because the alternative is a $400-an-hour lawyer they can't justify for a $3,000 project. Today's kit is a Claude-powered contract scanner that closes that gap, at 19 dollars a month and 10 to 30 cents of tokens per scan.

Let's get into it. Nine repos in Drops, three signals worth your attention, and the full ClauseGuard build below for those who have Upgraded.

Operator Access is a paid tier. Get the repo, the schema, the prompts, and the full clause library at theaigent.xyz/upgrade.

[Repo] anthropics/skills, The official Anthropic repo for Agent Skills, 146,506 stars. Build a reusable ability once and invoke it with /name across every Claude Code session. If you run repeating workflows and you're not using skills yet, this is the repo to clone first.

[Repo] anthropics/claude-code, The agentic coding tool that lives in your terminal, 130,124 stars. Understands your codebase, executes routine tasks, and handles complex explanations inline. The backbone of the operator stack this newsletter covers.

[Repo] RonitSachdev/ccundo, Granular undo for Claude Code, 1,401 stars. Reads directly from Claude Code's session files to track every file operation and lets you roll back precisely, without a full revert. The gotcha Claude Code doesn't have out of the box, handled.

[Repo] tripleyak/SkillForge, Intelligent skill router and creator for Claude Code and Codex, 681 stars. Analyzes any input to recommend existing skills, improve them, or generate new ones from scratch. Useful when your skill library starts getting large enough to lose track of.

[Repo] uppinote20/claude-dashboard, A status line plugin for Claude Code, 481 stars. Tracks context usage, API rate limits, and cost in real time. If you've ever been surprised by a compaction mid-session or hit a rate ceiling during a long run, this surfaces it before it costs you.

[Repo] himself65/finance-skills, A collection of Claude Code skills for AI financial analysis and trading, 2,696 stars. Ready-made skills you can invoke for financial reasoning tasks without building the prompt chain yourself. Good starting point before you write your own from scratch.

[Repo] QwenLM/qwen-code, An open-source terminal-based AI coding agent, 24,912 stars. Runs locally, lives in your shell like Claude Code does. Worth testing for cost-sensitive runs or projects where you want a second model in the mix.

[Repo] huggingface/smolagents, A minimal library for agents that reason in code, 27,706 stars. No abstraction overhead, no framework lock-in. When you want to wire a quick multi-step agent without a full framework, this is the baseline.

[Repo] ashishpatel26/500-AI-Agents-Projects, A curated collection of 500 AI agent use cases with links to open-source implementations, 31,677 stars. More reference library than runnable project, but useful when you're scoping what's already been built before starting from scratch.

[MCP] modelcontextprotocol/servers, The official reference list of MCP servers from Anthropic. Every MCP build starts here: canonical implementations for GitHub, filesystems, databases, and more. If you're wiring Claude to a real tool and not starting from this list, you're doing it the slow way.

[MCP] Patdolitse/piia-engram, Local-first, MCP-compatible persistent memory across AI tools, Apache 2.0. One memory layer shared across Claude, your editor, and any other MCP-compatible agent. The thing you want when context evaporates between sessions and you're tired of re-explaining the same project state.

[MCP] EliasOulkadi/shokunin, 62 agent skills for Claude Code, Cursor, Windsurf, and OpenCode, with ChromaDB memory, built-in MCP servers, and declarative self-updates. Multi-model, open source, zero cost. If you work across more than one coding agent and want a shared skill layer, this is the closest thing to a cross-tool standard that exists right now.

Vercel updated its terms of service to address agentic workflows. The specific change acknowledges that developers now routinely grant AI tools direct infrastructure access and that platforms themselves act autonomously. If you deploy agents on Vercel, the new terms are worth reading, not because they restrict anything obvious today, but because this is the first major platform writing agentic behavior into the legal contract layer. Source: Vercel Changelog.

Uber hit its 2026 AI tools budget in four months and capped Claude Code usage across engineering. Simon Willison covered the math: at scale, uncapped AI tool usage becomes a line-item problem fast. The lesson for solo operators is the opposite of what Uber is dealing with: your cost structure at 1-10 users is a fraction of what Uber pays per seat, and the capability is the same. The gap is closing from both directions. Source: Simon Willison.

Qwen-Code shipped as a terminal-native open-source coding agent with 24,912 stars at launch. It lives in the shell the same way Claude Code does. The practical question for any operator running Claude Code daily is whether a second agent in the mix, one running a cheaper open-weights model for lower-stakes tasks, changes the economics of a session. The answer is probably yes for the right workloads. Source: QwenLM/qwen-code.

This week: Hooks. Auto-run a command on any Claude Code event before or after it happens.

Most operators learn slash commands and CLAUDE.md early. Hooks are the move that comes after: you define a trigger (a tool call, a file edit, a bash execution) and attach a command that runs automatically when that event fires. No manual step, no reminder, no post-session cleanup.

Three practical setups:

1. Lint on every file edit. Add a PostToolUse hook for the Edit tool that runs your linter. Claude edits a file; the lint check runs immediately; broken syntax never sits. 2. Block dangerous bash before it runs. Add a PreToolUse hook on Bash that checks the command string against a deny list (e.g. rm -rf, DROP TABLE). The hook exits non-zero and Claude sees the rejection before execution. 3. Log every tool call. A PostToolUse hook that appends tool_name, input, and a timestamp to a local file gives you a full session audit trail without touching the Claude Code UI.

Hook config lives in .claude/settings.json under the hooks key. Each entry names the event, the matcher (which tool or command), and the shell command to run.

You'll know it's working when you edit a file and see your linter output appear in the Claude Code terminal before Claude's next turn.

Claude Code docs: hooks

Move: Git-checkpoint before every Claude Code run.

Claude Code is aggressive about file edits. If you're running it on a real codebase and something goes wrong mid-session, ccundo helps, but a clean git checkpoint is faster and more reliable.

1. Before any non-trivial Claude Code session, run git stash or git commit -am "pre-claude checkpoint [date]". 2. Let Claude run. Let it edit freely. 3. When the session ends, git diff HEAD shows you exactly what changed. If you want none of it, git reset --hard HEAD puts you back in under a second.

The payoff: you stop being nervous about letting Claude run multi-step edits. The safety net is in git, not in hoping Claude didn't touch something it shouldn't have. You'll know it's working when you stop manually reading every edit before approving it, and just let the diff tell you at the end.