Opening

Photo-editing software photocraft shot up to 39.3k stars on GitHub within a month, making it the most talked-about project in the open source community this week. But what actually won engineers over has little to do with photo editing, I'll explain why in a moment. We'll also talk about how to get an AI agent to draw an arrow on your screen and call in a human hand-off, plus an open source TPU whose design involved an AI agent, down to generating the circuitry itself, with real benchmark numbers laid out for you to see.

This Week's Open Source Picks

1. photocraft: A Photo Editor Shell That's Really a Command Interface for AI Agents

• What it is: An image editor written in Rust, but its real design core isn't about clicking through edits with a mouse. The whole editor is built as a command registry, complete with a CLI, a JSON control channel, and a built-in MCP server that lets agents drive the image-processing pipeline directly.
• This week's buzz: 39,275 GitHub stars (jumped to 39.3k ★ within 30 days), making it the loudest project on this week's open source radar. Licensed under Apache-2.0, written primarily in Rust.
• Highlights:
  • Written entirely in Rust, with a GPU compositor running on wgpu (Metal/Vulkan/DX12/WebGPU), no Electron or webview shell involved
  • PSD/PSB read/write passes 307 out of 309 test files, with over 1,700 tests across the whole project, and it even pits a CPU reference compositor against the GPU compositor as an oracle to cross-validate results
  • Can run entirely from the command line in batch mode, for example: photocraft-cli run wave.psd --cmd filter.sharpen.smartSharpen --params '{"amount":80}'
• Quick start: Download the 0.5.0 installer from getartcraft.com/apps/photocraft, free on all platforms with no account required. If you'd rather not install anything, grab photocraft-web-0.5.0.zip to self-host the WASM version and try it out. The source code is at cargo run --release -p photocraft -- image.psd.
• Compared to similar tools: It beats GIMP and Krita by reading PSD/PSB and Affinity files directly while preserving layers, and it comes with a native MCP/CLI automation interface out of the box. That said, the README itself flags it as early alpha, it's still missing around 20 tools, has no plugin compatibility, and some on HN questioned whether it's AI-generated and whether the hype outpaces the actual experience.
• Who it's for: Engineers who want an agent to take over the image pipeline directly. It's fun to play with as a tech demo, but hold off on putting it into production.
• Link: github.com/storytold/photocraft

2. big-arrow-on-the-screen (bigarrow): Letting AI Agents Draw an Arrow on Your Screen to Call You In

• What it is: A screen annotation tool built specifically for AI agents. When agents like Claude Code or Codex get stuck on a system permission dialog, 2FA, or CAPTCHA and need a human to click something, it draws an arrow or box on screen pointing to "click here manually," skipping the whole back-and-forth of the agent trying to describe a button's position in text.
• This week's buzz: 400 points on Hacker News (181 comments). 544 GitHub stars, MIT licensed, written primarily in Swift.
• Highlights:
  • bigarrow point --element "Allow" --app "System Settings" can target UI elements directly, or you can use --at 760,500 to point at coordinates; the markers disappear automatically after being drawn (8 seconds by default for point, 300 seconds for start), and none of it requires any macOS permissions
  • bigarrow install-skill installs the skill into ~/.claude/skills and ~/.agents/skills, costing just 182 tokens at rest and only consuming 1,528 tokens when triggered
  • Exit codes are designed for agents to interpret (0 for success, 2 for bad arguments, 3 for target not found, 4 for missing permissions), plus --copy-button, {{value}} for paste, --say for voice narration, and --follow to track a moving target; it can draw across multiple monitors and full-screen Spaces
• Quick start: brew install franzenzenhofer/tap/bigarrow, then run bigarrow install-skill once installed to hook it up to Claude Code/Codex. If you want to see it in action first, just bigarrow point --at 760,500 --text "點這裡".
• Compared to similar tools: It runs in the opposite direction from automation tools like cliclick or AppleScript, those click things for you, whereas bigarrow marks the spot for you when a real human has to step in.
• Who it's for: macOS users who run Claude Code/Codex daily and keep getting stuck on permission popups. The limitations are clear-cut too: it only supports macOS 14 and above, and building it yourself requires Xcode 16+.
• Link: github.com/franzenzenhofer/big-arrow-on-the-screen

3. OpenTPU: An Open Source TPU Co-Designed by an AI Agent, No Hardware Required to Run the Simulator

• What it is: A complete, full-stack open source TPU implementation, open sourced from the SystemVerilog RTL and custom ISA to a bit-accurate Python simulator, a kernel language, and a compiler. Even without an FPGA card, you can use the simulator to actually run LLM inference and see how the hardware was designed.
• This week's buzz: 346 points on Hacker News (402 comments), the most-discussed post of the week. 589 GitHub stars, Apache-2.0 licensed, written primarily in Verilog.
• Highlights:
  • Actually runs ten models, including Qwen3, Qwen3.5, Gemma 4, SmolLM3, and Phi-4-mini, on a Xilinx Kintex-7 xc7k480t (an Inspur YPCB-00338 PCIe card with two DDR3 channels, peaking at 17.1 GB/s)
  • The benchmark numbers get specific: LFM2.5-230M (4-bit) decodes at 85.8 tok/s with prefill at 335.4 tok/s; Qwen3-0.6B (int8) hits 21.6 tok/s; Qwen3.5-2B (4-bit) reaches 12.09 tok/s, and every single output matches the simulator token for token
  • The architecture deliberately hides nothing: no cache, no hidden scheduling, every single data move is its own instruction, so a single trace shows exactly where the cycles go. It also ships host-side tools: otpu-chat, otpu-smi, and the otpu-lens profiler
• Quick start: No need to buy a card. Clone the repo, run pip install -e ., then run otpu-chat --model lfm2 --backend isa, and you can chat with it locally using the simulator.
• Compared to similar projects: The name is easy to confuse with Google's academic OpenTPU (a TPUv1 replica), but this one is a full-stack implementation that can actually run modern LLMs. Compared to software-level optimization projects like llama.cpp or tinygrad, this one goes all the way down to the RTL and ISA layer.
• Who it's for: Hardware-minded engineers who want to understand how AI chips are designed, or who actually have an FPGA card and want to run it for real.
• Link: github.com/FeSens/openTPU

Closing Thoughts

These three projects share something in common: open source is no longer just software built for humans to use, it's increasingly handing control directly over to AI agents. Whether it's photocraft's MCP/CLI, bigarrow turning "ask a human for help" into a command an agent can call, or OpenTPU letting an agent take part in hardware design itself. The one I'd recommend engineers try first is bigarrow: one line of brew install, and five minutes later it's plugged into your agent workflow. Same time next week, the open source picks newsletter will dig up more new finds for you.