agent-look
Finds the screenshots you just took, reads the text via Spotlight OCR, renames generic files, and reports back with the useful evidence. It is the scout with binoculars.
Spottingagent-look starts with a visible identity: a practical little scout that spots your screenshots, reads the trail, and brings back the useful part. The bigger idea is a family of tools, skills, and plugins that each know one job cold, then fade into the background as agents start doing the job directly or learn exactly when to use the tool for you.
We are not building a pile of vague abstractions. We are building a family of practical tools that agents can trust. Each one is legible at the beginning, useful on day one, and progressively less visible as the agent stack absorbs the pattern.
Finds the screenshots you just took, reads the text via Spotlight OCR, renames generic files, and reports back with the useful evidence. It is the scout with binoculars.
SpottingPackage repeatable moves into named capabilities. A good skill starts visible enough to teach the pattern, then becomes something the agent reaches for automatically at the right moment.
Learned routinesExtend the agent into real tools and environments. Browser control, local state, external systems, file access: practical reach with a clear job boundary and a clean handoff back to the agent.
ReachGive the tool a crisp purpose and an identity strong enough that people understand when it should be used.
One command, one habit, one obvious move. The tool should solve a real annoyance immediately, not after a migration project.
As the agent stack matures, the skill becomes implicit. The capability remains, but the user no longer has to think about the machinery.
The scout metaphor is only useful if the tool is dependable. These implementation choices keep it light, practical, and boring in the right way.
It uses mdfind -onlyin … kMDItemIsScreenCapture=1 so iCloud-backed screenshot folders still work even when a direct filesystem walk hits macOS permission walls.
Spotlight OCR via kMDItemTextContent handles most screenshots for free. Vision fallback is reserved for the cases where OCR is empty, so the tool stays fast and cheap.
macOS screenshot names can include tricky spacing characters. fs.renameSync() with exact Spotlight paths avoids the fragile shell quoting path and just renames the file correctly.
agent-look register wires into Claude Code, Claude Desktop, Gemini CLI, and Codex. That keeps the tool recognizable now while positioning it to become part of the default agent workflow later.
The install story should match the product story: direct, practical, no ceremony. Homebrew gets the tool onto the machine, then agent-look register wires it into the agent surfaces you already use.
brew tap ashrocket/agent-look
brew install agent-look
agent-look register
/plugin marketplace add ashrocket/agent-look
/plugin install agent-look@agent-look
git clone git@github.com:ashrocket/agent-look.git
cd agent-look && bash install.sh
Use agent-look status to inspect registrations and agent-look unregister to remove them cleanly.
The package includes the MCP server, the screenshot examiner agent, and the skill wiring so the scout is available in the tools that support it.
File an issue if registration or screenshot discovery breaks. The goal is for this to feel invisible, so friction is a bug.
`agent-look` is one member of a broader pattern. `agent-pb` handles the last useful artifact like a clipboard knife. Browser plugins like `superpowers-chrome` extend the same practical philosophy into live inspection and control. Different jobs, same standard: strong identity now, invisible usefulness later.