SlimSnap: Giving Your CLI Agents the Eyes They Always Needed
The Screenshot Problem in AI-Assisted Development
If you've been using CLI-based coding agents like Claude Code, Aider, or Codex CLI, you've probably hit a frustrating wall: these tools are exceptional at reading and writing code, but they can't see your UI. Want to describe a broken button? You end up typing paragraphs like "the red error message in the top-right corner of the checkout page." Your agent squints at your words and tries to imagine what you mean.
Enter SlimSnap—a nifty macOS utility that transforms any screenshot into structured JSON that your CLI agent can actually process.
How It Works: Capture, Annotate, Ship
The workflow is beautifully simple:
Capture - Hit ⌘⇧S, drag to select any area of your screen. It uses native macOS APIs, so no additional dependencies.
Annotate - Drop arrows, callouts, and highlights directly on the screenshot to point out exactly what needs fixing.
Copy as JSON - One click converts everything to structured JSON, ready to paste into any terminal agent.
The JSON output is surprisingly rich. It includes the captured timestamp, app name, screen dimensions, and—most importantly—a structured array of detected elements. Each element comes with its type (button, input, label), actual text content, and bounding box coordinates in normalized 0-1 format.
This last part is crucial. When your agent knows the exact position of "the third input field in the settings modal," it can reason precisely about what to change instead of guessing.
Why This Matters for Developer Workflows
Here's where SlimSnap genuinely shines: token efficiency.
When you paste a raw screenshot into ChatGPT's web interface, the API downsamples and compresses it. But with CLI agents that accept images, you're still hitting API caps—around 1,568 vision tokens for a typical screenshot on Sonnet models, and up to 4,784 on Opus 4.7/4.8.
SlimSnap's JSON export typically lands between 600-800 tokens. That's roughly 55% fewer tokens per interaction on Sonnet, climbing to 85% savings on Opus models. Over a long debugging or refactoring session with dozens of iterations, those savings compound quickly.
More tokens left in your context window means more room for actual code, error logs, and agent reasoning. Your context budget stretches further.
Privacy-First Architecture
One thing I appreciate: SlimSnap does all OCR processing locally on your Mac. Captures never leave your machine. No account required, no server in the loop. If you're working with sensitive applications or proprietary interfaces, this matters.
The JSON schema itself is open (MIT license) and published on GitHub. You can validate against it, write your own exporters, or even hand-craft valid JSON if you want to integrate with other OCR pipelines.
The Claude Code Skill
For Claude Code users specifically, SlimSnap offers a skill that automatically detects your captures. It writes a small config file to ~/.slimsnap/config.json containing your save preferences, and the skill reads that to load the latest JSON into your agent's context. Seamless integration without any manual file hunting.
When JSON Beats Pixels
Here's the honest trade-off: for element-level work—fixing specific buttons, updating form fields, adjusting layout issues—JSON is objectively better. Your agent gets precise coordinates and extracted text instead of interpreting pixels.
But for "vibe checks"—does this design feel cohesive? Does the color palette work?—you might still want to paste the raw image. Nothing prevents you from sending both: the JSON for precise targeting, the image for aesthetic reasoning.
The Bottom Line
SlimSnap solves a specific but real pain point in the AI-assisted development workflow. If you're deep into CLI-based agents and frequently need to discuss UIs, it's worth adding to your toolkit. The token savings alone justify the setup, and the local processing means zero privacy concerns.
Currently macOS only, but since the schema is open, it's only a matter of time before someone builds compatible exporters for other platforms.
Check it out at slimsnap.ai—the productivity gains for UI-heavy debugging sessions could be substantial.
Tired of describing UIs in paragraphs? Sometimes the best tool is the one that translates what you see into what your agent understands.