AI Drops: Amazon Knockoff Filters, App-Generating Search, Local AI, and Workflow Memory

Direct Answer

This week's most useful releases are not all the most ambitious ones. Knockoff removes a daily shopping frustration. Maker Skills gives agents reusable operating methods. Printing Press turns websites and APIs into more efficient agent interfaces. Osaurus makes local models approachable on Apple Silicon.

The higher-trust tools are more interesting and more demanding. HeyClicky can read screen context and voice. social can act on LinkedIn and X. Screenpipe can build searchable memory from work activity. Their value comes from access, so the test plan must begin with permissions, exclusions, and visible approval gates.

Video credit: The Next New Thing, with Andrew Warner and Corey Ganim. Watch the original channel on YouTube. The episode says it is presented by Zapier; this article is independent and not sponsored by Zapier or the featured products.

Source Note

The video and supplied transcript provide the discovery layer and the hosts' hands-on reactions. Product capabilities, operating-system support, privacy behavior, access models, repository links, and pricing status were checked against the official product sites, privacy pages, and GitHub repositories on 16 July 2026.

The word free in the episode title needs qualification. JustVibe currently offers free generated apps without sign-up. Maker Skills, Printing Press, and Osaurus publish open-source code. SceneRoll offers a free trial before paid plans. social meters upstream usage. PlugThis is subscription software. Screenpipe combines an open-source local core with commercial product layers. Availability and prices can change.

One creator attribution also needs care. The episode discusses an infinite-canvas demo associated with Meng To. The source project that matches the demonstrated Codex workflow is Cowart, maintained under the zhongerxin GitHub account. This article credits the demo signal and the repository separately instead of assuming they are the same author.

Link Map

What I Would Test First

1. Shopping Filters and Context-Aware Input

Knockoff: Make Amazon Search Less Hostile

Knockoff is a Chrome, Firefox, and Safari extension by Josh Pigford that filters suspicious Amazon brands, can hide sponsored listings, and lets the user maintain personal trust and block lists. Its three levels range from relaxed filtering to strict recognized-brand-only results. The site says searches stay on-device and that network calls are documented and individually controllable.

That privacy posture is good, but the product is deliberately not absolute. Knockoff says its verdicts use heuristics plus community lists and can be corrected. Use it to reduce search noise, then still inspect seller identity, warranty, return path, reviews outside Amazon, and safety certification for batteries, chargers, health products, children's products, or anything that can cause harm.

HeyClicky: Dictation Becomes a Screen-Aware Assistant

Farza's HeyClicky sits near the cursor, sees the current screen, accepts voice instructions, and can launch background agent work. The demo drafts an email reply using the visible message and a learned voice profile. That is more capable than literal speech-to-text: it combines speech, screen context, memory, and action.

It is also not a fully local dictation tool. The official privacy policy says push-to-talk screenshots and voice input are captured locally, then sent through HeyClicky's backend proxy to Anthropic, OpenAI, AssemblyAI, and ElevenLabs. Avoid testing it while password managers, private client data, health information, financial accounts, unreleased designs, or confidential chats are visible. For fully local basic dictation, compare it with FluidVoice.

2. Search Results and Browser Extensions Become Apps

JustVibe: An Interactive Answer Instead of Ten Blue Links

JustVibe describes itself as a search engine that answers with a free interactive app. A travel request can become a planner with routes, activities, and adjustable preferences instead of a paragraph. The episode's Reykjavik demo is honest about the current limit: the first result was useful, but some controls were awkward and the one-shot app needed refinement.

Treat app-generating search as a better interface for organizing evidence, not as evidence itself. Verify opening hours, prices, routes, safety advice, medical claims, and financial calculations against primary sources. The app should show source URLs and observed dates before it earns trust.

PlugThis: A Focused Builder for Chrome Extensions

PlugThis generates Manifest V3 Chrome-extension code from plain English, can add a Supabase backend, supports bring-your-own model keys, and lets the user download and own the code. That specialization can remove setup friction around content scripts, background workers, manifests, packaging, and store submission.

The security review still belongs to the owner. Check every requested host permission, content-script match pattern, data store, API key path, external request, analytics call, update URL, and user-facing disclosure. Start with a no-backend extension that needs only activeTab and local storage. Do not use generated code to scrape accounts or automate actions that violate a site's rules.

3. Agents Need Better Interfaces Than Raw Browser Clicking

social: LinkedIn and X as Structured Commands

The social CLI exposes profiles, posts, comments, reactions, followers, messages, connections, local sync, and SQL queries to agents. Structured JSON and local mirrors can be more reliable and token-efficient than repeatedly asking a browser agent to click through a social network.

The main caveat is authority. The official site says login defaults to read,write unless write access is cleared. Write commands include posts, comments, reactions, follows, direct messages, and LinkedIn connection requests. Start with read-only scope, a test account where possible, no scheduled outreach, and human approval for every external action. Residential proxying and rate limiting do not remove your responsibility to follow platform terms, outreach law, employer policy, and recipient consent expectations.

Cowart: A Project-Local Visual Canvas for Codex

Cowart is an open-source infinite-canvas plugin built around tldraw. It lets Codex open a visual workspace, read selected canvas context, place generated images, and iterate from annotations while keeping project assets local. That is useful for mood boards, design references, image holders, and visual feedback that is clumsy to describe in prose.

The right first test is one disposable project with non-sensitive assets. Inspect the repository, install scripts, plugin manifest, local server, MCP tools, and write paths before giving it a real client workspace. Keep generated images traceable to their prompts and source assets.

Printing Press: Teach the Agent a Stable Command Surface

Matt Van Horn's Printing Press can generate an agent-native Go CLI, companion skills, and MCP server from an API or website. Its library installs community CLIs alongside agent instructions. The design idea is strong: a local mirror and one compound command can replace repeated remote calls and brittle multi-step browser work.

The self-learning update described in the episode lets the generated tool record lessons and reduce future calls. That memory must remain reviewable. A CLI that silently adapts to a private endpoint can preserve a mistake as easily as it preserves a successful fix. Begin with read-only queries, pin versions, inspect diffs, test against a non-production account, and keep website-derived CLIs under maintenance because private endpoints change.

4. Creator Workflows Become Reusable Systems

SceneRoll: Record B-Roll Once, Reuse It

SceneRoll turns audio or unedited vertical video into transcript-based scenes, then matches each scene with Pexels stock footage or clips from the user's own vault. The vault is the valuable part: clips are tagged and reused, so the fifth video can be faster than the first. Its current official plans include a no-card trial and paid creator tiers.

Test it with one 30-second video and ten owned B-roll clips. Measure edit time, caption corrections, wrong clip matches, pacing changes, and final export quality. Verify commercial rights for every stock, music, image, and uploaded asset before publishing.

Maker Skills: Free Operating Methods for Agents

Corey Haines' Maker Skills repository focuses on the personal operator's craft: decisions, research, second-brain work, scenario modeling, content, finance workflows, video analysis, and meta-skills for authoring more skills. It works across Claude Code, Codex, and Cursor and is published under the MIT license.

Do not install the entire library merely because it is free. Pick one recurring task, read its SKILL.md and scripts, inspect external calls and write behavior, run it in a sample workspace, and compare the result with your existing prompt. Keep it only if it reduces correction time or improves a measurable output.

5. Osaurus Makes Local AI Feel Like a Normal Mac App

Osaurus is a native Swift app for Apple Silicon that can connect Ollama, MLX, or LM Studio, run offline, read files, use tools, and add cloud providers when a task exceeds the local model. The official site says it requires macOS 15.5 or later, needs no account, includes no telemetry for local use, and is MIT licensed.

The episode highlights model-fit guidance, which solves a real beginner problem: knowing whether a model is too large, tight, or comfortable for the machine. Local does not mean harmless, though. A local agent with file writes, browser plugins, folder watchers, and execution rights can still delete or leak data. Begin offline with a small model, one read-only folder, no shell execution, and an acceptance test that checks factual quality, latency, memory use, and refusal of forbidden actions.

6. Screenpipe Turns Work Activity Into Searchable Memory

Screenpipe records screen and audio activity on-device, makes it searchable, and can turn patterns into SOPs and agents. The official product shows grounded questions across meetings, apps, and handoffs; background agents with visible triggers and actions; and connections to common work tools. Its open-source repository makes the recording layer inspectable.

Its privacy controls are more concrete than the video alone suggests. Screenpipe says excluded apps, windows, and URLs are dropped before reaching disk, while an on-device model can scrub cards, Social Security numbers, and keys from remaining captures. Those controls reduce risk; they do not solve governance by themselves.

Pilot it with one volunteer, one machine, seven days of retention, no personal apps, no password manager, no banking, no health data, no private messaging, and no automatic external actions. Review what was captured before enabling search, SOP generation, or team-wide memory.

The Larger Pattern: Context Is Moving Closer to the Work

These launches form one coherent stack:

1. Filter the environment: Knockoff removes low-trust results before the user acts.
2. Understand the current screen: HeyClicky turns visible context and speech into an instruction.
3. Return an interface, not prose: JustVibe and PlugThis generate something executable.
4. Give agents structured tools: social, Cowart, and Printing Press replace fragile generic clicking.
5. Accumulate reusable assets: SceneRoll grows a B-roll vault; Maker Skills preserve methods.
6. Move inference local: Osaurus makes open models easier to own.
7. Turn history into memory: Screenpipe tries to learn from the work itself.

The design rule is the same at every layer: the closer a tool gets to real context and real actions, the more explicit its boundaries must become. Better context improves results. Narrow permissions, evidence, and approval gates keep those results governable.

A Seven-Day Test Plan

1. Day 1: choose one tool for one repeated pain. Record the current time, cost, and error rate.
2. Day 2: read its pricing, privacy policy, permissions, repository or export model, and deletion path.
3. Day 3: run the smallest test with synthetic or non-sensitive data.
4. Day 4: test one failure: bad source, wrong clip, broken API, ambiguous brand, or denied action.
5. Day 5: measure accepted output after human correction, not the prettiness of the first demo.
6. Day 6: remove one unnecessary permission and confirm the workflow still works.
7. Day 7: keep, limit, or uninstall it. Document the decision and the next review date.

Bottom Line

The most practical tools in this episode are Knockoff for a narrow consumer problem, Maker Skills and Printing Press for reusable agent infrastructure, SceneRoll for a clear creator workflow, and Osaurus for approachable local AI on modern Macs. JustVibe, Cowart, and PlugThis are promising examples of interfaces generated around the task instead of forcing the user into a generic chatbox.

HeyClicky, social, and Screenpipe may unlock more value, but they also touch the most sensitive surfaces: what is on the screen, who the user contacts, and what the company remembers. Test them last, start read-only, and make every consequential action visible to a person.

Sources

• The Next New Thing: Free - Stop Amazon knockoffs, vibecode apps, local AI, and more
• The Next New Thing, Andrew Warner, and Corey Ganim
• Knockoff
• HeyClicky, privacy policy, and Farza
• JustVibe
• social CLI
• Meng To and zhongerxin/Cowart
• SceneRoll
• Printing Press, CLI repository, and public library
• Corey Haines: Maker Skills
• PlugThis
• Osaurus and GitHub repository
• Screenpipe and open-source repository