🛠️ 9 Claude Code Tools Worth Cloning This Week

Three things landed today that change your cost math as an operator.

GPT-5.6 hit general availability this morning in three sizes: Luna, Terra, and Sol. Simon Willison has the token pricing breakdown. The short version: the frontier got cheaper at the bottom and more capable at the top, and that gap is where you make decisions about which model routes which work.

Anthropic is moving in the other direction. Wired reports that Claude subscribers will soon pay usage-based fees on top of their flat subscription to access Fable 5. The flat-rate era is ending. Budget your agent calls accordingly.

The most practically useful thing I found today is not a model at all. It is OfficeCLI, a single-binary, no-Office-required toolkit that lets agents read, edit, and automate Word, Excel, and PowerPoint files. 13,268 stars and it ships as one binary. If any part of your pipeline touches documents a client actually uses, clone this today.

[Repo] iOfficeAI/OfficeCLI, 13,268 stars. A single binary that gives AI agents full read/edit/automate access to Word, Excel, and PowerPoint with no Office installation required. If your pipeline ever hands off to a client who lives in Office files, this is the missing link.

[Repo] MadsLorentzen/ai-job-search, 18,450 stars. An AI-powered job application framework built on Claude Code. Fork it, fill in your profile, and it evaluates postings, tailors CVs, writes cover letters, and preps you for interviews. Good architecture to steal for any personalization-at-scale workflow.

[Repo] open-webui/open-webui, Self-hosted ChatGPT-style UI for Ollama and OpenAI-compatible endpoints. The go-to front end when you want a clean interface over local or private models without touching the API directly.

[Skill] rohitg00/skillkit, 1,360 stars. Install, translate, and share skills across Claude Code, Cursor, Codex, Copilot, and 40+ more coding agents from one place. If you build a skill once and want it portable, this is the layer that handles the translation.

[Repo] 0xranx/OpenContext, 655 stars. A personal context store for AI agents and assistants with a desktop GUI to capture, search, and reuse context across sessions. Works with Claude, Codex, and OpenCode out of the box.

[Repo] djyde/ccmate, 624 stars. Configures Claude Code without the pain. If you have ever wrestled with settings files across projects, this is the tool that makes configuration feel sane.

[Repo] disler/infinite-agentic-loop, 602 stars. An experimental two-prompt system that demonstrates an infinite agentic loop in Claude Code. Good reference architecture for any workflow that needs to run autonomously without a human poking it.

[Skill] forayconsulting/gemini_cli_skill, 187 stars. A Claude Code skill that routes calls to Gemini 3 Pro via Gemini CLI. Useful when you want to compare outputs or use Gemini's context window on a specific task without leaving your Claude Code session.

[Repo] Helicone/helicone, 5,928 stars. One line of code to add observability to any LLM call: monitor, evaluate, and experiment. If you are running agent pipelines and have no visibility into what they are spending or where they fail, this is the fastest fix.

[MCP] Zhenyu98/codex-chatgpt-bridge, A bridge that lets Codex and ChatGPT hand off coding work to each other, saves tokens, and keeps local execution under your control. The non-obvious use: route expensive reasoning steps to whichever model is cheapest for that task class and keep the rest local.

[MCP] Orkas-AI/Orkas-VideoStudio, Describe a video in plain language and your coding agent writes the timeline and produces the file. If you are building any content pipeline that needs video output, this is the fastest way to wire it into an existing agent workflow.

GPT-5.6 is generally available in three sizes. Luna, Terra, and Sol are live for any company or individual today. The operator read: you now have three new price points to slot into your routing logic.

Claude Fable 5 may go usage-based on top of subscription. Flat-rate AI subscriptions are ending at the frontier. Wired reports Anthropic will charge extra for Fable 5 access on top of current monthly fees. Audit which of your agent calls actually need the best model and route the rest down.

Bun rewrote itself in Rust in 11 days for $165K in tokens. Bun's team documented the full rewrite for teams considering AI-assisted rewrites. The real signal is not the speed. It is that $165K in tokens did what a small team would have taken a year to do. That arithmetic is going to show up in more roadmap conversations this quarter.

Muse Spark 1.1 is on Vercel's AI Gateway. Meta's multimodal reasoning model with a 1M token context window now routes through Vercel's gateway. Accepts text, image, video, PDF, and audio. If you are already on Vercel, you can swap it in as a drop-in for long-context multimodal tasks today.

This week's technique: MCP servers. Not the setup. The part most operators skip: scoping them correctly and confirming they actually connected.

Most operators install an MCP server and assume it worked. They find out it did not when Claude hallucinates a tool call three steps into a task.

Here is the discipline:

1. Scope it to the project, not the user. Drop your MCP config into .mcp.json at the repo root. That config travels with the repo. User-scoped servers (in ~/.claude.json) are for tools you want everywhere; project-scoped is for tools that only make sense in this context. Mixing them is how you get a database MCP firing in a repo where it has no business being. 2. Confirm the connection before you build on it. Open a session in that repo and run /mcp. You will see every server Claude Code has loaded and its status. If a server is missing or errored, you catch it now, not mid-task. 3. Name your tools explicitly in the first prompt. If you are relying on the GitHub MCP to file issues, say "use the GitHub MCP to create an issue" rather than just "create an issue." Claude will pick the right tool. Without the explicit call, it may reach for a web search or a bash command instead and produce a mess you have to unwind.

You will know it worked when /mcp shows green status and your first tool call completes without a fallback.

Claude Code docs: MCP

The move: wire Helicone into your agent pipeline in five minutes and find out what it is actually spending.

Most operators are flying blind on token cost per task. You know the monthly bill. You do not know which agent call is burning 40% of it.

1. Add one line: swap your openai.OpenAI() (or Anthropic client) base URL to https://oai.hconeai.com/v1 and pass your Helicone API key as a header. That is the entire integration. 2. Run your normal workflow for 24 hours without changing anything else. 3. Open the Helicone dashboard and sort by cost. You will see which prompts, which tasks, and which models are generating most of the spend.

Cut the top offender first, whether that is a prompt that is sending 10k tokens of context when 2k would do, or a task class that is hitting GPT-5.6 Sol when Luna would be fine.

You will know it worked when you see a per-call cost breakdown you can act on, not a monthly invoice you can only wince at.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: the revenue step that failed quietly for sixteen days.

A revenue integration can break in a way that makes every dashboard happier, not sadder. That is what makes it dangerous.

Our signup page runs a recommendation step after someone subscribes: a partner network surfaces other newsletters, and when a reader opts in, that is a revenue event for us. During a page rework, a post-submit redirect got added that skipped the step entirely. Signups kept flowing. The page looked cleaner than ever. Conversion on the form actually reads better when you remove steps. Nothing errored, because nothing was broken in the part the monitors watched.

It ran sixteen days before a live walkthrough caught it. Sixteen days of every ad-fed signup skipping the one step that pays for the ads.

The lesson is not "test more." We had tests. The lesson is that green dashboards measure what you told them to measure, and nobody had told anything to measure the money step. Presence of traffic is not presence of revenue.

The fix beyond the fix: a canary. A synthetic signup runs on a schedule and asserts one thing only, that the revenue step actually fired. Not that the page loaded, not that the form submitted. That the step that pays happened. If it did not, that is the alert. You are not alerting on errors anymore; you are alerting on the absence of a heartbeat.

The copyable pattern: list every step in your funnel that creates money. For each one, answer "what tells me within a day if this silently stops." If the answer is a human noticing, build the canary this week.

Got a step in your funnel that could go quiet without a single error? Hit reply and tell us what your canary would check. We read every one.

Nine repos, two skills, two MCPs, and four signals worth acting on. The one I keep coming back to is OfficeCLI: agents that can finally touch the files clients actually send you is a bigger unlock than it sounds. Clone it this weekend and see what breaks.

See you Monday.