Five fresh agent, video, advertising, and ComfyUI projects, plus five durable open-source classics.
B.Y.E. / DAILY AI TRY / 10.31.26

Be Your Evolution AI News Daily

October 31, 2026 · Five newer tools explored, five proven classics worth keeping

Top 5: New and Worth Your Attention

Agnix repository
TRY NOW

1. Agnix 0.60.2

Agnix is a linter and language server for agent configuration: AGENTS.md, SKILL.md, hooks, MCP settings, permissions, and files used by Codex, Claude Code, Cursor, OpenCode, Copilot, and others. Its browser playground makes the first test unusually low friction. Jason could paste a non-sensitive sample configuration, inspect the diagnostics, then run the CLI against a disposable repository to catch ignored fields or unsafe defaults before an agent silently behaves differently than intended.

OPERATIONAL VALUE
Lint one copied agent setup, review every finding against the tool’s evidence, and record which warnings reveal real configuration drift versus stylistic preference.

QUICK VIDEO
Plant three subtle errors in an AGENTS.md and MCP config, run Agnix, then show which mistakes it catches and which still require human review.

Reality: The project is dual-licensed MIT or Apache-2.0 and offers npm, pip, Cargo, Homebrew, binaries, and a no-install playground. Version 0.60.2 was released October 30. Its reported rule count and editor coverage are upstream claims; a clean scan is not a security audit. Avoid pasting secrets into the playground, pin the CLI version, inspect autofixes before accepting them, and independently validate permission and MCP behavior.

OPEN PRIMARY SOURCE →
WeftCut repository
TRY NOW

2. WeftCut 0.5.6

WeftCut is a desktop nonlinear editor whose timeline can be driven through a localhost MCP server while every edit remains visible and manually editable. Agents can place, trim, split, caption, keyframe, mix, checkpoint, and create code-based overlays; the review view exposes activity and before/after timeline differences. Jason could duplicate a short interview project, ask an agent to mark pauses before cutting, and approve or revert the proposed edits instead of surrendering the timeline.

OPERATIONAL VALUE
Use copied media, request review markers first, inspect the timeline diff, then export a short test and compare duration, captions, audio sync, and source preservation.

QUICK VIDEO
Let an agent tighten a 30-second interview while playback runs, then reveal the exact added, removed, and changed regions before approving the edit.

Reality: WeftCut code is MIT and v0.5.6 was released today for Windows, macOS, and Linux. The app bundles no model weights; speech, vision, cloud providers, optional local engines, source footage, music, fonts, and exports retain separate terms. The project says its macOS installer is not yet verified. Keep MCP local, grant only copied project media, preserve originals, review every edit, and verify the exported file independently.

OPEN PRIMARY SOURCE →
LoomWatch repository
WORTH A LOOK

3. LoomWatch 0.21.0

LoomWatch is a Windows and macOS desktop workspace for building agent teams, routing handoffs, pausing for approval, and replaying a local timeline of work. Each agent can have its own model, skills, readable inputs, permissions, and deliverable. That makes it interesting for editorial experiments where a researcher hands evidence to a writer and a checker. Jason could test whether the run receipt makes omissions, costs, and human decisions clearer than copying answers between chats.

OPERATIONAL VALUE
Create a three-stage team using public material only, require approval between stages, then compare the local receipt with the final brief for lost citations or unsupported claims.

QUICK VIDEO
Watch a researcher, writer, and checker hand one story across a live timeline, pause the workflow, reject weak evidence, and resume with a corrected source.

Reality: The repository is Apache-2.0 and version 0.21.0 was released today. Model access is separate: the documented options include ChatGPT-plan access, Claude API credit, or OpenRouter, each with its own privacy, price, and terms. “Nothing else to install” describes the packaged app, not zero setup or zero cost. Start with public inputs, no filesystem writes, explicit approvals, low spending limits, and an independent check of the final receipt.

OPEN PRIMARY SOURCE →
adsctl repository
WATCHLIST

4. adsctl 0.11.0

adsctl is an agent-ready CLI for Meta advertising reports, monitoring, and recoverable campaign operations. It separates read scopes from write profiles, requires explicit account selection for writes, bounds pagination, and ships setup and analyst skills for compatible agents. The valuable first use is not launching an ad. Jason could run a read-only performance audit against a deliberately scoped test account and examine whether the structured output exposes incomplete pages, unavailable metrics, and ambiguous results clearly.

OPERATIONAL VALUE
Keep the profile read-only, use a test account, request one bounded report, and reconcile its IDs, date window, completeness flags, and totals against Meta’s own interface.

QUICK VIDEO
Show an agent attempting an unsafe write, the CLI refusing it, then switch to a read-only audit and verify the same numbers in Meta Ads Manager.

Reality: adsctl code is MIT, but Meta’s proprietary APIs, accounts, tokens, advertising policies, billing, audience data, and campaign content remain separate. Release 0.11.0 provides Windows binaries, yet persistent launch and monitoring are explicitly unsupported on Windows. Setup requires a Meta developer app and Marketing API permissions. Do not pipe an installer blindly, expose tokens to prompts, enable write mode, or spend money during an evaluation.

OPEN PRIMARY SOURCE →
GOAT'd Text Generator repository
WORTH A LOOK

5. GOAT’d Text Generator for ComfyUI

GOAT’d Text Generator adds prompt, caption, summary, and description generation inside ComfyUI. It can use hosted providers, Ollama, LM Studio, or a generation-capable model loaded in ComfyUI, and it can inspect images plus sampled video frames. Jason could connect a local model to a copied workflow, turn a rough visual idea into structured prompt variants, and compare whether the node improves repeatability without sending media to a cloud endpoint.

OPERATIONAL VALUE
Import the local example workflow, replace optional viewer nodes if needed, use non-sensitive media, and log the model, sampling settings, prompt, timing, and final text.

QUICK VIDEO
Feed one image and silent video sample into the local node, generate three model-specific prompts, then compare the resulting images beside the original rough idea.

Reality: The node code is GPL-3.0 and has no tagged release, so pin a reviewed commit. It requires a recent ComfyUI V3 node API; models, Ollama or LM Studio, optional GGUF loaders, and provider keys are separate. Frame sampling omits soundtrack content, and media support varies by model. Review the node’s dependencies, keep keys in environment variables, prefer loopback local servers, and verify every generated prompt and output license.

OPEN PRIMARY SOURCE →

5 Proven Classics

MICROSOFT AUTOGEN · WATCHLIST

AutoGen remains an instructive multi-agent Python framework with AgentChat, Core, Studio, and MCP support, but Microsoft now identifies Agent Framework as its production-ready successor. Use AutoGen deliberately, not by default.

Source →
AGNO · WORTH A LOOK

Agno remains an Apache-2.0 platform for agents, teams, workflows, APIs, MCP, approvals, and durable execution. Disable optional telemetry if desired, and start with one constrained service before adopting its broader platform.

Source →
SMOLAGENTS · TRY NOW

Hugging Face’s Apache-2.0 smolagents stays useful for compact code-agent experiments across local and hosted models. Use Docker or another sandbox for generated code, constrain tools, and inspect every action.

Source →
CHROMA · TRY NOW

Chroma remains approachable Apache-2.0 search infrastructure with Python and JavaScript clients. Prototype locally, then test deletion, backup, tenant isolation, embedding drift, and retrieval quality before storing sensitive knowledge.

Source →
TABBY · WORTH A LOOK

Tabby remains a capable self-hosted coding assistant with Docker deployment and local model options. Hardware, model licenses, repository access, editor extensions, authentication, and update cadence still need separate evaluation.

Source →

Reality Check

A repository license covers code, not the whole workflow. Agnix can flag known configuration mistakes but cannot prove an agent is safe. WeftCut keeps edits visible, yet footage, music, models, local engines, and exports retain separate rights. LoomWatch still relies on paid or separately governed model access. adsctl touches regulated advertising accounts and money, so read-only scope and human approval are mandatory. GOAT’d can stay local, but its models, loaders, ComfyUI build, and generated media each need review. Pin versions or commits, use copied data, isolate credentials, cap spending, preserve originals, and independently verify every consequential result.

Bottom Line

Try Agnix first because its browser playground and disposable-repository scan can reveal configuration drift without connecting a production account. Then give WeftCut a copied 30-second interview and require review markers before edits; its visible timeline diff directly supports Jason’s preference for editable, auditable creative handoffs. LoomWatch is the next comparison for multi-agent evidence. Keep adsctl read-only, and treat GOAT’d as a local ComfyUI experiment until its untagged commit and dependencies have been reviewed.

Curated from official project repositories and release notes. Verify current terms before production use; software, services, models, data, identities, media, accounts, advertising, and outputs may carry different licenses.