11 JC Lab Ideas to Film, Test and Write Today: Agent Sounds, Quota Forensics and Persistent Codex (2026)

Start with Beckon, tare, or yolobox. Each one gives you a visible result in a short session: a distinct agent-state sound, a plain-English quota diagnosis from local logs, or a YOLO coding agent that cannot touch your real home directory. The rest of the pack adds four tool reviews, two crypto attention checks, and one article on OpenAI's still-unreleased Persistent Codex mode.

Video topics to film

Beckon: can four sounds replace staring at Claude Code?

Format: Tutorial

Shoot: Install with cargo install beckon-cli, run beckon init and film the settings diff plus backup path, then beckon test so the pack plays on camera. Trigger four states in Claude Code if you can: finished turn, needs a decision, failed, and rate-limited. Beckon's repository describes a no-daemon hook tool for Claude Code that plays a rising chime, insistent sting, falling tone, or slow pulse; it claims ~5 ms decisions, no network at hook time, no telemetry, unconditional exit 0, and dual MIT/Apache-2.0 licensing with CC0 built-in packs. Linux audio is fully verified; macOS and Windows build in CI but the README marks their audio as not yet heard. 1 2

Hook / verdict frame: "Did the sound tell me what to do, or was it still just a ding?" End on a four-cell card: finished / needs you / failed / throttled.

Check before publishing: Keep beckon init on a disposable Claude settings file. Say on camera that only Claude Code is supported today.

tare: where did the Claude quota actually go?

Format: How-I-tested

Shoot: Install with npx skills add kelviq/tare -g -y --copy --agent claude-code, open a fresh Claude Code session, and ask three questions in order: what burned tokens yesterday, which sessions look automated, and what to change first. Film the cause card, the evidence from local session logs, and one wrong total that tare corrects by deduplicating repeats. tare's repository says it audits Claude Code usage from logs already on the machine, tracks the rolling 5-hour window, recognizes automation patterns, and needs no account; MIT. 3 4

Hook / verdict frame: "Was the limit my work, or a background loop I never watched?" Put the top consumer next to the total.

Check before publishing: Use only your own local logs. Do not paste another person's session files into a public video.

yolobox: YOLO agent mode that leaves your home directory at home

Format: How-I-tested

Shoot: Install with brew install finbarr/tap/yolobox or the official install script, open a disposable project folder, start one supported agent in yolobox, and run a harmless command that would be scary outside a sandbox (rm -rf ~/something-fake rewritten so it only exists inside the box). Film the container mount, the missing host home, and the README's own safety limits. yolobox's repository describes a containerized YOLO wrapper for Claude Code, Codex, Gemini, OpenCode, Copilot, Pi, and similar CLIs: the project mounts at its real path, the host home stays out unless you opt in, and the author calls it protection from accidents rather than a container-escape theorem; MIT. 5

Hook / verdict frame: "Did YOLO feel fast because the sandbox was real, or because I still trusted the agent with the project?" Show the home-directory miss and the project mount side by side.

Check before publishing: Keep secrets out of the mounted project. Say clearly that hostile code still needs stronger isolation than yolobox.

Opslane: can session replay find a bug that never threw?

Format: How-I-tested

Shoot: Run docker compose up -d --wait, hit http://localhost:8082/health, then install the browser SDK on a tiny disposable frontend with a known dead-click or self-closing dropdown. Film one rage-click replay, the user-impact ranking, and the MCP path if your coding agent can ask what broke. Opslane's repository describes an AGPL-3.0 self-hosted system that records sessions, ranks issues by users affected, investigates with sandbox verification, opens a PR only when a fix verifies, and ships an MCP server; browser/Python SDKs are MIT, with input masking on by default. Full investigation needs Anthropic, E2B, and GitHub credentials. 6 7

Hook / verdict frame: "Did the agent catch a silent UX failure, or only the exceptions I already knew?" Cut from the replay to the issue card.

Check before publishing: Use synthetic users and a throwaway app. Treat auto-opened PRs as drafts until you review them.

Tools and apps to review

screenpipe: always-on computer history for agents, with a privacy scorecard

Format: Scam check

Shoot: Install the desktop app or CLI, exclude password managers and banking windows first, record 10 minutes of synthetic work, then ask an MCP client one question that only the capture can answer. Film Settings → Privacy, whether analytics is on by default, and what happens when you force local transcription. screenpipe's site and repository describe continuous local capture of screen, audio, and accessibility context for macOS, Windows, and Linux, with a local API and MCP server; the GitHub README marks the source as a commercial license for commercial use, notes PostHog analytics and Sentry on by default in the desktop app, and says cloud transcription or sync will leave the machine if you enable them. Product Hunt is running a launch-day BUSINESS20 annual discount through 28 August 11:59pm PT. 8 9 10

Hook / verdict frame: "Is this local history I can audit, or a recorder that phones home until I tighten every toggle?" End with a checklist: excluded apps, analytics off, local STT, MCP query that worked.

Check before publishing: Never film real client calls, passwords, or banking screens. Re-check the license before calling the project open source.

Firecrawl Developer Index: one search over docs, READMEs, issues, and PRs

Format: Review

Shoot: Run npx -y firecrawl-cli@latest setup developer-index, then search the same coding question three ways: plain web search, developer category search, and firecrawl developer "...". Film one issue hit, one README hit, and whether a keyless call still returns. Firecrawl's docs describe /v2/search/developer plus categories: ["developer"] on /v2/search, filters for docs/issues/PRs/READMEs/repos/stars, keyless access with a free allowance, and MCP/CLI surfaces; Product Hunt launch copy claims a curated index of 70M+ GitHub artifacts with no API key needed to start. 11 12 13

Hook / verdict frame: "Did the developer index beat generic search on the same question, or only feel denser?" Keep the three result sets on one card.

Check before publishing: Treat the 70M figure as Product Hunt launch copy. Prefer the docs for endpoint shape and filters.

Glisio: local Mac demo recorder with auto-zoom and free watermarked export

Format: Review

Shoot: On macOS 13+, record a 60-second demo of any throwaway app with system audio and mic, let auto-zoom follow clicks, then export the same take as 16:9, 9:16, and 1:1. Film the watermark on Free versus the Pro remove-watermark path if you trial it. Glisio's site lists Free forever with unlimited recordings and a watermark, Pro Lifetime at $79, Pro Monthly at $9.99, local offline edit/export, ScreenCaptureKit capture, and no BlackHole requirement. 14

Hook / verdict frame: "Is Free enough for YouTube drafts, or does the watermark force Pro on day one?" Show all three aspect exports from one recording.

Check before publishing: Re-check live pricing before stating a final number. Keep exports on your Mac if privacy is part of the claim.

Revalvo: run one prompt across every model, score it, keep a receipt

Format: How-I-tested

Shoot: Open revalvo.com with no account, connect one cloud key and one local Ollama or LM Studio model, paste the same prompt, and run them in parallel. Turn on two structure scorers and one latency/cost scorer, save a version, then rerun against a tiny CSV dataset. Revalvo's site describes a free local-first browser workbench with keys in IndexedDB, 40 evaluators, immutable scored snapshots, dataset batching, GitHub YAML sync, and provider support that includes OpenRouter, OpenAI, Anthropic, Groq, Ollama, and LM Studio. 15 16

Hook / verdict frame: "Did the scoreboard change which model I would ship, or only confirm the one I already liked?" Put pass rate, latency, and cost on one report frame.

Check before publishing: Use disposable keys and synthetic prompts. Treat LLM-judge scores as another model opinion, not ground truth.

Pons / PONS: can a mid-pack search spike clear a liquidity check?

Format: Trend react

Shoot: CoinGecko's 24-hour trending return placed Pons at score 10 with market-cap rank 269. A morning simple-price snapshot recorded about $0.137. CoinGecko page meta at fetch time also carried roughly $17.4 million in 24-hour volume. 17 Refresh while filming, then record live price, volume, spread, depth, and the timestamp from one liquid venue.

Hook / verdict frame: "Did attention arrive with a book you can measure, or only a rank you can screenshot?" Apply one written paper rule and include estimated slippage.

Check before publishing: The trend list measures search attention. Use paper trades only and label every market number with the capture time.

Ribbita by Virtuals / TIBBIR: same paper rule, different book

Format: Comparison

Shoot: CoinGecko's 24-hour trending return included Ribbita by Virtuals and gave it market-cap rank 137. A morning simple-price snapshot recorded about $0.291. Page meta at fetch time also carried roughly $3.1 million in 24-hour volume. 18 Refresh both cards while filming, then put TIBBIR beside PONS under the same paper rule, window, and exit condition.

Hook / verdict frame: "Does the higher-ranked name still lose once spread and depth enter the sheet?" Keep the rule identical so the only variable is the book.

Check before publishing: Ranks and prices move during the day. Re-capture both screens immediately before filming and avoid a buy, sell, or price target.

Article idea

OpenAI's Persistent Codex mode: keep working until put to sleep

Format: Article

Write: Open with WIRED's 27 August 2026 report that OpenAI is testing a Persistent mode for Codex. Walk readers through what the public Codex code and an OpenAI spokesperson confirm: the agent is instructed to continue until put to sleep, a related proactivity path can create follow-up tasks and message sparingly across sessions, Persistent mode does not expand permissions on its own, and there are no immediate launch plans. Tie the piece to the same week's research context WIRED flags around persistent-model risks and sandbox probing when agents face impossible tasks. 19

Reader promise: Separate three creator questions: what Persistent mode changes about session length, what it still cannot do without approval, and which logging or kill-switch habits a small lab should practice before any always-on agent ships. Close with a four-point checklist: hard spend caps, out-of-band tool logs, explicit sleep/stop commands, and no production credentials in a long-running agent host.

Check before publishing: Quote WIRED and the attributed OpenAI confirmation, not social summaries. Do not frame the piece as a product launch or as proof that consumer chatbots will act without limits.

Film in this order

Capture Pons and Ribbita first while the CoinGecko numbers are fresh; write the paper rule before you open either chart.
Install Beckon next: init, test pack, and four state sounds are a short A-roll block.
Run tare in a fresh Claude Code session while the logs from yesterday are still on disk.
Review Glisio and Revalvo in one sitting: both are browser- or Mac-first and need little setup.
Hit Firecrawl developer search with the same question three ways while the CLI is warm.
Build yolobox on a disposable project, then film one scary-looking command that stays inside the box.
Save screenpipe for a quiet room after you lock exclusions and turn analytics off.
Bring Opslane up with Docker only after the lighter installs are in the can.
Write the Persistent Codex article after you have opened the WIRED piece, copied the dated confirmation, and drafted the four-point kill-switch checklist.