10 AI Things Worth Trying
Five fresh signals. Five durable tools. No hype tax.
Top 5: New and Worth Your Attention
1. scrollcraft
scrollcraft is a Claude Code skill for building scroll-driven sites with more discipline than a generic “make it cinematic” prompt. It treats motion, typography, spacing, contrast, and mobile behavior as one system, then uses browser checks to catch dead scroll, weak contrast, broken opacity, and failed video. Jason’s practical experiment is one B.Y.E. campaign page: compare the skill’s plan, signature interaction, and contact sheet with an ordinary agent-built landing page.
OPERATIONAL VALUE
Reusable art-direction constraints plus automated visual checks for a branded landing-page workflow.
QUICK VIDEO
Can one skill turn a flat B.Y.E. brief into a scroll-directed launch page? Show the feeling curve, signature interaction, and contact sheet.
Reality: The documented install is a Claude Code marketplace command; native Codex compatibility is not claimed. Rich motion still needs assets and performance/accessibility review. MIT code; media rights remain separate.
OPEN PRIMARY SOURCE →2. agenttrail
agenttrail turns an agent’s declared plan and actual file edits into a live project map. That makes it easier to see whether a coding run is progressing, stuck, or reopening work it marked complete. The value is supervision without reading an entire transcript. Jason could point it at a small repository, start a Codex task, walk away, then use the board to identify one mismatch between stated intent and observed edits.
OPERATIONAL VALUE
Local, visual oversight for long coding-agent runs; its README says no account, global install, or telemetry is required.
QUICK VIDEO
Start npx agenttrail, dispatch a small Codex edit, and watch file activity move across the component map.
Reality: The richer map needs agenttrail init, PLAN.md, and agent-maintained conventions. Claude has richer hook/todo integration than Codex. MIT code.
3. Doop desktop v0.1.2
Doop gives people and agents the same live design canvas. Frames render real HTML in sandboxed iframes, while an MCP client can create and revise designs beside a human cursor. That is more inspectable than a prompt-and-refresh loop and could suit rapid campaign exploration. Jason’s useful test is to have an agent draft three B.Y.E. hero directions on one canvas, pin the strongest example, and revise it live.
OPERATIONAL VALUE
Agent-assisted visual iteration with shared context, comments, and editable HTML rather than flattened screenshots.
QUICK VIDEO
Connect an MCP client, request three campaign frames, then leave a pinned comment and watch the agent revise one live.
Reality: Local setup requires Git, Bun, and Node; Docker Compose is the documented production option. Agents need separate provider access. AGPL-3.0 repository; hosted and provider terms remain separate.
OPEN PRIMARY SOURCE →4. ACRYL v0.1.15
ACRYL tries to make the workspace, task history, artifacts, and handoffs belong to the project instead of one coding agent. Claude Code, Codex, OpenCode, Gemini CLI, and others can enter and leave the same development scene. The useful idea is continuity: switch agents without rebuilding context from scratch. Jason could run one small task in the TUI, hand it to a second agent, and inspect what context actually survives.
OPERATIONAL VALUE
A persistent local coordination layer for comparing or changing coding agents while retaining project state.
QUICK VIDEO
Begin a task with one harness, switch workers, and show the second agent inheriting the same artifacts and task state.
Reality: v0.1.15 explicitly remains early development. Desktop, global CLI, and local web are separate surfaces; Windows x64 has an installer. MIT code; model subscriptions and API usage remain separate.
OPEN PRIMARY SOURCE →5. Lemmalog
Lemmalog explores a different answer to agent memory: store facts and rules in a deductive database, preserve where facts came from, and let the agent ask why a conclusion exists. That is more auditable than simply retrieving similar text. It is promising for long-running research or operations agents, but still young. Jason should first test a tiny set of changing project facts and inspect its proof tree and retraction behavior.
OPERATIONAL VALUE
Explainable, update-aware agent memory where conclusions can point back to source episodes.
QUICK VIDEO
Add a fact and rule, query the derived answer, display why(), then retract the source fact and watch the conclusion disappear.
Reality: This is a Rust engine, not a polished memory app; MCP is feature-gated and model-based extraction remains an ingestion boundary. Reproduce benchmarks before production use. MIT code.
OPEN PRIMARY SOURCE →5 Proven Classics
Run supported pretrained models in a browser or Node without a separate inference server; useful for private, low-friction demos. Apache-2.0 library code; individual model licenses still govern weights.
Source →A broad, actively maintained collection of vision backbones plus training, evaluation, inference, and export tooling. Apache-2.0 code does not make every linked pretrained checkpoint unrestricted.
Source →Mature Python NLP components for tokenization, tagging, parsing, entity recognition, and production pipelines remain useful when deterministic structure beats another generative-model call. MIT code; pipeline packages have separate metadata.
Source →Still useful for topic modeling, document similarity, and compact vector-space baselines when a transparent classical method is preferable to an LLM. LGPL-2.1 library obligations deserve review before distribution.
Source →The reference speech-recognition and translation repository remains a dependable baseline for turning interviews or field recordings into text. MIT code; model weights, source audio, privacy, and consent need separate review.
Source →Reality Check
“Open source” here refers to repository code under the named license. Model weights, hosted services, provider APIs, user data, supplied assets, and generated-media rights remain separate. scrollcraft is Claude Code-specific; agenttrail’s Codex view is file observation plus conventions. Doop and ACRYL are young experiments. Lemmalog is an engineering prototype whose production fitness and performance claims need independent testing.
Bottom Line
Start with agenttrail: it can make existing coding-agent work easier to supervise through a small, reversible local trial. Try scrollcraft second when a visually strong B.Y.E. landing-page demo is the priority. Keep Lemmalog on the watchlist.
Primary-source research completed August 31, 2026. Private B.Y.E. field guide.