11 JC Lab Ideas to Film, Test and Write Today: Desktop Agents, Local Tools and the Crypto Trend Trap (2026)

Hands and FetchSandbox deserve the first two recording slots: one puts a real desktop-control boundary on camera, while the other gives you a planted API bug and a visible receipt to compare with a green test run. The crypto ideas belong at the end because the numbers can move before the upload; record them on screen when you shoot, then keep the conclusion at the level of research and paper trading.

Video topics to film

Hands: can a desktop agent stop itself?

Format: Scam check

Shoot: Use a spare Windows profile and a public, non-sensitive page. Hands describes itself as a Windows eyes-and-hands Model Context Protocol (MCP) and command-line tool that can see the desktop and move the real mouse and keyboard in Chrome. Its README also says the tool is not a sandbox, warns about prompt injection and money or account actions, and names Pause/Break as the kill switch. 1 Give it one harmless navigation task, add a page instruction that tries to redirect the task, and press Pause/Break during a second run. Keep payments, personal accounts, live accounts, and real money off the test machine.

Hook / verdict frame: Show the cursor, the task prompt, the page's competing instruction, and the exact moment the kill switch stops control. The verdict is about one bounded safety test: whether the agent obeys a visible stop boundary under controlled conditions.

Check before publishing: Treat Hands as untrusted desktop control. Keep the browser profile disposable, explain that a public-page test cannot certify safety, and show every permission or access step the audience would need.

Dictata: local dictation with cleanup on or off

Format: Comparison

Shoot: Dictata's repository describes 100% local, system-wide Windows dictation, with Whisper.cpp as the default backend, optional ONNX speech backends, and an optional local OpenAI-compatible large-language-model cleanup step. The current README marks Linux as unsupported. 2 Read the same 45-second JC Lab script three times: default Whisper.cpp, another available speech backend, and cleanup turned on versus off. Record start-to-first-word latency, transcription errors, cleanup changes, and CPU or memory use on the same computer.

Hook / verdict frame: Which setting gives a creator the best balance between private processing, speed, and edit-ready text? Put the raw transcript and cleaned transcript side by side, then show the clock instead of repeating a generic "local AI" promise.

Check before publishing: Confirm which backends and model files are available on the machine that day. Separate speech-recognition errors from LLM rewriting, and label hardware results as one-machine measurements.

FetchSandbox MCP: does a green test prove the integration works?

Format: How-I-tested

Shoot: The official FetchSandbox playground contains five brownfield applications with planted API integration bugs. Its demo repositories are designed to run without real API keys. 3 Clone one app, run its tests before changing anything, point an agent at the documented integration gap, and capture both the test output and the sandbox's actual request or response receipt. Use fake data and keep production credentials out of the session.

Hook / verdict frame: Put two screens beside each other: the green test suite and the observable integration result. The video should answer whether the test checks the behavior that a user actually cares about, rather than treating a passing test as the whole verdict.

Check before publishing: Keep the claim narrow. A planted bug in one playground app can demonstrate how the workflow behaves on that fixture; it cannot establish broad API reliability or prove the product's launch-copy coverage claim.

LayoutLens: build three broken interfaces and measure them

Format: Tutorial

Shoot: LayoutLens documents a deterministic layer built around axe-core WCAG checks plus geometry, contrast, and occlusion checks. It also has an optional vision-LLM tier, while the deterministic findings remain able to overrule the model. 4 Create three small UI defects: unreadable contrast, an occluded button, and a misplaced element. Run the deterministic checks first, enable the vision tier only if access is available, and compare the findings in a simple result log.

Hook / verdict frame: Can a vision model find a visual problem that a measured rule misses, and what does the model miss that geometry catches? The finished tutorial should leave viewers with a repeatable test, not a score to admire.

Check before publishing: Keep deterministic failures and model suggestions in separate columns. State whether the optional model call used an API key, and do not present the vision tier as a replacement for measured accessibility checks.

Tools and apps to review

Construct Computer: what does an AI coworker keep?

Format: Review

Shoot: Construct describes its service as a cloud-based work platform with AI agents. Its privacy policy says the platform stores workspace files, chat and session history, agent memory, workflows, scheduled jobs, calendar data, activity history, access-control records, and connection state or credentials for integrations. 5 If access is available, create a dummy workspace, give the agent one harmless research task, and inspect the workspace, memory, workflow, scheduled-job, and integration screens afterward. Use dummy files and dummy identities.

Hook / verdict frame: After one task, can you find what the agent retained, what it can reach, and how you would remove it? The review earns its conclusion by showing the storage and permission surfaces rather than by repeating the word "workforce."

Check before publishing: Do not connect real Slack, Telegram, Discord, Google, payment, or personal accounts. Record the access scope, deletion controls, retention language, and any waitlist or plan limit you encounter. Treat the privacy policy as the declared data practice and the product test as a check of the visible user controls.

OpenLogi: a local Logitech control panel in practice

Format: Review

Shoot: OpenLogi's site calls the app a Rust, local-first alternative to Logitech Options+, using HID++ for Logitech devices. The site says it has no account and no telemetry, 44 built-in actions, TOML-owned bindings, and signed macOS, Linux, and Windows builds; per-app profiles are marked as coming later. 6 Use a compatible mouse, remap two buttons, change DPI, switch SmartShift, edit one TOML binding, and reopen the app to test persistence. Disconnect the network for a second run if the device still works.

Hook / verdict frame: Does the local workflow give you the controls you actually use, and what breaks when a feature depends on Logitech's wider device ecosystem? Show the config file, the physical result, and the missing per-app profile in the same review.

Check before publishing: Treat "no telemetry" as the developer's stated claim, then report what your network observation can and cannot establish. Test pairing, receiver import, device model, and platform separately; the site says it is not affiliated with Logitech.

Postlia: one JC Lab post across several accounts

Format: Review

Shoot: Postlia's channels page lists LinkedIn, Bluesky, Instagram, TikTok, Pinterest, YouTube Shorts, and Mastodon publishing, plus a 7-day free trial. 7 Take one short JC Lab clip, prepare it for two supported accounts, and record the work from upload to scheduled preview. Compare the caption, title, privacy setting, thumbnail or image treatment, and approval step for each destination. Publish only after checking every field.

Hook / verdict frame: Does one source clip become two platform-ready posts, or does the scheduler leave the important editing to you? Put the original caption beside the two previews and time the setup.

Check before publishing: Use a test post or an account with a clear rollback path. Confirm the account permissions and the exact trial end date in the live interface; a 7-day trial says nothing about the long-term price or whether every feature remains available after it.

Claude Academy: turn one lesson into a working demo

Format: How-I-tested

Shoot: Anthropic's official collection lists Claude Code 101 as 12 lessons and one hour, Claude Platform 101 as 13 lessons and 1.5 hours, and courses covering skills, subagents, MCP, and the Claude API. 8 Complete one short course or the first useful module, then build a small JC Lab demo from the lesson: for example, a local content brief generator or a tool-calling mockup. Time the lesson, setup, first useful output, and one correction.

Hook / verdict frame: How long does it take to move from a lesson to a working result? Show the starting task, the first failed or incomplete attempt, and the final demo so the audience can judge learning value from the work produced.

Check before publishing: Treat the catalog duration as a course estimate, not a promise about your learning speed. Verify access, model costs, and any course completion or certification language inside the current course before mentioning it.

SPX6900 and PIPEDOG: attention versus tradability

Format: Trend react

Shoot: The current CoinGecko trend return surfaced SPX6900 and PIPEDOG for the reaction shortlist. Use their CoinGecko pages to identify the assets, then record the live price, market-cap rank, 24-hour volume, spread, and order-book depth at the moment of filming. 9 10 Run a paper entry and exit with a fixed small notional, include estimated slippage and fees, and timestamp the screen.

Hook / verdict frame: Does search attention produce a tradable market, or only a noisy headline? Put attention, liquidity, spread, and simulated execution in one frame. The reaction ends with a research question, never a buy or sell conclusion.

Check before publishing: CoinGecko trend membership is a popularity signal. Volatile values need a fresh screen capture, and a paper trade says nothing about what a real order would fill like later.

Ethena / ENA: read the yield before reacting to the trend

Format: Trend react

Shoot: Ethena's site presents USDe and sUSDe, currently displays a lifetime-average sUSDe APY of 10.7%, and says sUSDe is not a bank deposit or insured product. 11 The current CoinGecko trend return also surfaced Ethena. 12 Build the reaction around four screens: what USDe is, what sUSDe adds, what the displayed rate measures, and what backing, redemption, custody, transparency, and insurance questions remain.

Hook / verdict frame: What exactly is a viewer being shown when a dashboard says "APY"? Put the advertised rate beside the product's own assumptions and risk disclosures, then paper-track the token separately from the yield product.

Check before publishing: Label 10.7% as Ethena's current site display and lifetime-average claim, not a guaranteed return. Keep the video educational and paper-based; do not turn the rate, trend membership, or token price into an investment conclusion.

Article idea

no-telemetry: how much privacy can 43 tools prove?

Format: How-I-tested

Write: The no-telemetry README describes a CLI with zero production dependencies, no runtime network requests, a curated registry, and a scan of one project's direct dependencies. The current registry lists 43 tools: 34 environment-variable opt-outs, 2 opt-in checks, and 7 config-only tools. The commands include npx no-telemetry init --dry-run, npx no-telemetry doctor, and npx no-telemetry check; the README says the tool does not scan workspaces or transitive dependencies. 13 Run the CLI on a small JavaScript project, save the JSON or terminal output, then test one known dependency and one package outside the registry.

Hook / verdict frame: Can a narrow registry make a useful first check without pretending to be a complete privacy audit? The article should show the command, the result, the missing transitive case, and the exact question the reader must investigate next.

Check before publishing: Keep the 43-tool figure tied to the current README and record the commit or version shown during testing. Explain the difference between a direct-dependency check and a workspace-wide or transitive audit; avoid turning a clean result into a claim that the project is private.

Film in this order

Record Hands first while the spare Windows profile and safety boundary are ready.
Run FetchSandbox while the planted bug and test receipt are easy to reproduce.
Capture Dictata and LayoutLens before moving to the longer app reviews.
Review Construct, OpenLogi, Postlia, and Claude Academy with dummy data and visible settings.
Record the two crypto reactions last, with live screens and paper trades.
Write the no-telemetry piece from the saved CLI output and its narrow scope.