Five practical AI projects, five durable classics, and the setup and license reality behind each.
B.Y.E. / DAILY AI TRY / 10.20.26

Be Your Evolution AI News Daily

October 20, 2026 · Five newer tools explored, five proven classics worth keeping

Top 5: New and Worth Your Attention

AgentRelay Portable interface
WATCHLIST

1. AgentRelay Portable 1.1.0

AgentRelay Portable moves selected conversations and skill files among Codex, Claude Code, WorkBuddy, DeepSeek Harness, CodeBuddy, and supported JSON exports. Version 1.1.0 adds WorkBuddy AI separation, Tencent Yuanbao JSON import, English controls, and stronger cross-device packaging. The universal bundle includes runtimes for Windows, macOS, and Linux, so a useful test is exporting one non-sensitive session, previewing the conversion, and checking whether its messages, tool results, and skill draft survive intact.

OPERATIONAL VALUE
Create an inspectable, portable archive of approved agent work and reduce dependence on one client’s local history format before moving between machines or tools.

QUICK VIDEO
Export one harmless Codex conversation, preview its normalized record, restore it on a second folder or machine, and compare messages, tool results, titles, and skill files.

Reality: The code is MIT open source and the release publishes checksums plus a GitHub build attestation. Support is uneven: the project distinguishes file conversion, native-package restoration, client indexing, and actual conversation continuation, and some WorkBuddy AI desktop behaviors remain unverified. The bundle is unsigned on macOS and handles sensitive local histories, so work from copies, inspect the preview, exclude secrets, verify the checksum, and never treat a successful file move as proof that the destination client can safely resume the conversation.

OPEN PRIMARY SOURCE →
Launch Video Kit brand preview
TRY NOW

2. Launch Video Kit

Launch Video Kit turns an owned website into an editable brand kit and a 30-to-60-second product film. Ten agent skills cover brand extraction, story, voice, score, sound, motion, mixing, review, and visual world-building. The film’s copy and timing live in JSON, while screenshots and brand assets remain replaceable. That makes it unusually aligned with reusable creative handoffs: an agent can build the first cut, but Jason still receives inspectable source, audio stages, review notes, and a normal MP4.

OPERATIONAL VALUE
Convert an approved site into a repeatable launch-video package whose script, screenshots, brand tokens, timing, audio, and rendered output can be revised independently.

QUICK VIDEO
Extract a brand kit from one owned landing page, approve the preview sheet, replace the fictional script, then reveal the JSON timeline behind the finished 30-second film.

Reality: The repository is MIT open source, but the setup is not one click: its documented path uses Node.js, FFmpeg, uv, whisper.cpp, Python dependencies, and the HyperFrames renderer, with the current quick setup written for macOS. The included voice, alignment, music, and sound path is a starting point, not automatic creative approval. Use only owned or cleared websites, logos, screenshots, fonts, voices, music, and footage; verify third-party notices; and review every extracted claim before publishing.

OPEN PRIMARY SOURCE →
Tracekit tamper-evident agent tracing banner
TRY NOW

3. Tracekit 1.4.0

Tracekit is a tamper-evident flight recorder and policy gate for agent runs. It can instrument supported frameworks, inspect tool calls before execution, record model and tool activity into a signed hash chain, and verify whether the record changed later. Its local demo needs no API key and deliberately shows a destructive hidden instruction plus an edited log. That gives Jason a concrete way to test whether an agent workflow produces evidence strong enough to review after a mistake instead of trusting its own summary.

OPERATIONAL VALUE
Add reviewable policy decisions and integrity checks to selected agent workflows, then preserve an export that can reveal hidden tool calls or later log tampering.

QUICK VIDEO
Run the scripted cover-up demo, compare the agent’s cleaned log with Tracekit’s failed hash verification, then block one risky tool call with a visible policy rule.

Reality: Tracekit is MIT open source, but coverage varies by integration: its own compatibility table labels some paths stable and others experimental. A verified hash chain proves the captured record has not changed; it does not prove every relevant event was captured, that a policy was correct, or that the model’s output was truthful. Traces can contain prompts, tool arguments, paths, and sensitive data. Start with a disposable mock workflow, inspect storage and redaction behavior, and never treat observability as permission to grant broader access.

OPEN PRIMARY SOURCE →
Kokon local image editor
WORTH A LOOK

4. Kokon 0.9.4

Kokon packages local Qwen-Image 2.1 generation and photo editing into a desktop application with an MCP server and local HTTP API. Version 0.9.4 adds native Linux packages, macOS 15 support, reversible color-LUT versions, and faster inpainting. Windows, macOS, and Linux users can describe edits, restyle or relight images, and retain local project versions without uploading media. For Jason, the strongest test is agent-directed editing that still leaves every image version visible and the original untouched.

OPERATIONAL VALUE
Keep approved photos on the workstation while exposing controlled generation and editing to an agent, with visible versions that preserve the source and make comparisons straightforward.

QUICK VIDEO
Relight an owned photo, mask one object for inpainting, apply a neutral LUT, then ask an MCP-capable agent to create a second version without overwriting either result.

Reality: Kokon’s application code is AGPL-3.0 open source, while Qwen weights, GGUF conversions, the VAE, upscaler, and generated-media rights retain separate terms. The practical floor is substantial: Windows and Linux list roughly 12 GB VRAM, 16 GB RAM, and 15 GB free disk; an 8 GB GPU path is beta, as are AMD, Intel, Mac, and Linux support. macOS builds are not notarized. Verify release hashes, keep the server local, preserve originals, and test one cleared image before committing storage or production work.

OPEN PRIMARY SOURCE →
ComfyUI MOSS Prompt Enhancer repository preview
WATCHLIST

5. ComfyUI MOSS Prompt Enhancer

ComfyUI MOSS Prompt Enhancer generates MOSS-Image prompts locally through a GGUF language model. It supports caption and structured layout modes, preserves requested Chinese text, and can reverse-engineer a connected image into a reusable prompt when paired with a compatible vision projector. The node only writes prompts; it does not generate the final image. That separation is useful for repeatable art direction because Jason can inspect the prompt, spatial boxes, dimensions, token usage, and repair count before passing anything downstream.

OPERATIONAL VALUE
Turn rough art direction or an approved reference into inspectable prompt and layout data that can be saved, compared, corrected, and reused inside existing ComfyUI workflows.

QUICK VIDEO
Feed one cleared reference image into reverse mode, reveal the generated character and composition description, correct one layout box, then compare two MOSS-Image outputs from the same prompt.

Reality: The node code is Apache-2.0 open source, but it requires an existing ComfyUI installation, Python 3.10 or newer, a compatible llama.cpp server, and separately licensed GGUF weights; image mode also needs the matching visual projector. The author’s text-flow test used a large 27B quantized model, so VRAM, RAM, speed, and quality vary sharply. Its offline tests cover plumbing and safety checks, not real inference quality. Keep references cleared, verify model cards, align layout dimensions to multiples of 16, and judge every resulting prompt and image manually.

OPEN PRIMARY SOURCE →

5 Proven Classics

DAGSTER · WORTH A LOOK

Dagster remains a strong Apache-2.0 orchestrator for observable data assets and pipelines. Its local tooling aids debugging; deployment infrastructure, integrations, and Dagster+ commercial features remain separate choices.

Source →
LANCEDB · TRY NOW

LanceDB remains an Apache-2.0 embedded retrieval library for text, images, and other vectors. It suits local prototypes; embedding models, source data, cloud services, and production scaling require separate decisions.

Source →
RASA OPEN SOURCE · WATCHLIST

Rasa remains an Apache-2.0 framework for controlled text and voice assistants with explicit dialogue logic. The public repository is mature, but recent activity and commercial platform boundaries deserve review before adoption.

Source →
DEEPEVAL · TRY NOW

DeepEval remains an Apache-2.0 testing framework for LLM outputs, RAG, and agents. Its metrics create repeatable checks, but judge selection, evaluation data, model calls, and hosted Confident AI remain separate.

Source →
DSPY · WORTH A LOOK

DSPy remains an MIT framework for optimizing modular language-model programs against measurable goals. It can reduce manual prompt tweaking; datasets, evaluators, provider costs, and optimization overfitting still need discipline.

Source →

Reality Check

Open-source code is only one layer. AgentRelay can move files without proving a destination client can resume them. Tracekit can verify captured records without proving coverage or policy quality. Launch Video Kit still depends on a multi-tool media stack and cleared creative assets. Kokon and the ComfyUI node require separately licensed models and meaningful local hardware, while “local” servers still need access controls. Preserve originals, use disposable tests, verify hashes and model cards, minimize permissions and keys, keep sensitive traces private, and require human review before publishing any migrated history, agent action, prompt, image, voice, or video.

Bottom Line

Try Launch Video Kit first on an owned landing page because it produces the most reusable creative handoff: brand tokens, editable story JSON, staged audio, review artifacts, and a finished MP4. Run Tracekit’s no-key demo next; it is a compact way to test what trustworthy evidence around agent actions should look like. Keep AgentRelay, Kokon, and the MOSS Prompt Enhancer as focused lab trials until their compatibility, hardware, model terms, and output quality match a real workflow.

Curated from official project repositories and documentation. Verify current terms before production use; software, services, models, data, and content may carry different licenses.