Daily AI Implementation Scout Council

Agent Plugins 1.0.0 (agentplugins/agent-plugins-spec) repo

An open, vendor neutral specification, version 1.0.0, published August 6, 2026 by a steering committee spanning AWS, Anysphere (Cursor), Microsoft, OpenAI, and Vercel, that defines a single plugin folder format bundling Agent Skills and MCP server configuration so one package installs across competing agent products.

What it does for you: Gives YY a standard container to package Axion's Skills so they can run in Claude Code today and in ChatGPT, Cursor, Copilot, or Codex tomorrow without a rewrite, since a plugin is just a plugin.json manifest plus an optional skills folder and an optional mcp.json. It also gives Hermes a ready made packaging format to adopt instead of inventing its own, lowering the cost of ever sharing a skill outside the Skills repo.

In practice: The most consequential item found this week because it is not one project competing for attention, it is five rival vendors agreeing on the container format underneath all of them, which rarely happens this cleanly in this industry.

For: Both runtimes. It is a packaging specification, not tied to any single agent runtime, so both Axion's Claude based skills and a future Hermes runtime can package extensions against the same schema.

Verdict: build now. Score 31. Low execution risk since packaging one existing skill into the spec's folder layout is a reversible, local change, and the strategic upside for YY's 710 skill portfolio is large.

Source · 1.3k GitHub stars and 68 forks on the spec repo as of 2026-09-06, version 1.0.0 published August 6, 2026 with a version 1.1.0 working draft already in progress, governance charter caps any single company's Core Maintainer seats so no vendor holds a majority, confirmed by direct fetch.

agentgateway repo

An open source, self hosted proxy purpose built to sit in front of AI agents and MCP servers, described by its maintainers as a next generation agentic proxy, handling routing, policy, and traffic between agent clients and the tool servers behind them.

What it does for you: If Axion ever runs more than a handful of MCP servers at once, this is a single front door that could apply consistent routing and policy instead of YY hand wiring each server into Claude Code separately, and it is a candidate shape for the kind of approval or policy layer the GHL write gate hook already does by hand.

In practice: Solid infrastructure play rather than a flashy demo, the kind of piece that matters more as Axion's MCP server count grows than it does today.

For: Both runtimes. It is a standalone proxy binary that sits in front of any MCP server or agent, so it is not specific to Claude or to a future Hermes runtime.

Verdict: test first. Score 24. Useful direction but no immediate pain point it solves for YY's current MCP setup, worth a small trial before any real commitment.

Source · 4.7k GitHub stars and 808 forks as of 2026-09-06, confirmed by direct fetch.

langwatch repo

An open source, Apache 2.0 licensed platform for LLM evaluation and AI agent testing that instruments agent frameworks, runs real time or batch evaluations, and includes a separate scenario framework for scripted multi turn agent tests using an agent under test, a user simulator agent, and a judge agent.

What it does for you: Gives YY a ready made way to score whether a skill like hook writer v2 or speak marc is actually improving output instead of relying on eyeballing it, using the same simulated multi turn test pattern the hook writer v2 eval already tries to do informally.

In practice: A mature, well documented evaluation platform rather than a thin wrapper, with the scenario testing idea being the most directly reusable piece for YY's own skill evaluation habit.

For: Both runtimes. It is reached over an SDK and API, not tied to Claude specifically, so both Axion and a future Hermes runtime could send traces and evaluations to it.

Verdict: test first. Score 24. Directly relevant to YY's existing interest in evaluating skills against a weaker model, worth a small pilot on one skill before wider adoption.

Source · 3.5k GitHub stars and 364 forks as of 2026-09-06, Apache 2.0 license, confirmed by direct fetch.

Agent Squad (2FastLabs) repo

A flexible, lightweight multi agent orchestration framework, formerly named multi agent orchestrator and previously hosted at awslabs/agent-squad, now moved to 2FastLabs and extended this month with a Swift runtime for on device orchestration on iPhone, iPad, and Mac, running agents, MCP tools, realtime voice, and tracing entirely on device.

What it does for you: Not an immediate fit since Axion already uses Claude Code as its own orchestrator, but the on device Swift runtime is worth knowing about if YY ever wants an offline or mobile facing agent that does not round trip through a server.

In practice: A credible, actively maintained project that just changed owners and added a real new capability, not just a version bump.

For: Both runtimes. Ships as npm, PyPI, and now a Swift package, so it is not tied to any one agent runtime and could sit alongside either Axion or Hermes.

Verdict: watch. Score 23. No current on device or mobile need to justify adoption, but the ownership move and Swift runtime are worth tracking.

Source · 7.8k GitHub stars and 738 forks as of 2026-09-06, Apache 2.0 license, confirmed move from awslabs/agent-squad, confirmed by direct fetch.

Microsoft Agent Framework repo

Microsoft's converged agent framework, formerly two separate AutoGen and Semantic Kernel projects, shipped version 1.17.0 this month with a Foundry hosted Telegram sample, broader OpenAI SDK 3.x support, a Mistral SDK migration, and fixes to approvals, history replay, and hosting reliability.

What it does for you: Mostly reference value for YY rather than a direct build, since Axion is Claude based and does not run on Microsoft Foundry, but the approvals and history replay patterns are worth studying since Axion's own GHL write gate and night shift queue solve a similar human approval problem.

In practice: A large, well funded, actively shipping framework, moving fast enough that last month's review is already several versions behind.

For: Both runtimes. Available in both .NET and Python and framework agnostic to which model answers, so it fits either a Claude based approach or a separate Hermes runtime.

Verdict: watch. Score 23. Different stack from Axion's Claude based approach, tracked for its approval and replay patterns rather than for direct adoption.

Source · Version 1.17.0 shipped within the past week per the Microsoft Agent Framework devblog, confirmed by direct search, prior review on 2026-08-30 was several versions earlier.

grok-build (xai-org) repo

xAI's coding agent harness and terminal UI, launched as a closed beta in May 2026, open sourced under Apache 2.0 in mid July, and reaching version 1.0 on August 7, 2026, built as a fullscreen, mouse interactive, extensible CLI in the same space as Claude Code, Codex CLI, and Gemini CLI.

What it does for you: Not a tool YY would run instead of Claude Code, but the first party notice that it vendors in source ports of openai/codex and sst/opencode tool implementations is a useful signal of how fast the coding CLI space is converging on the same tool call conventions.

In practice: A credible, well backed competitor rather than a toy, but not one that displaces YY's existing Claude Code workflow.

For: Standalone tool. It is a competing, closed ecosystem coding agent CLI from xAI, a full standalone product rather than something Axion would embed inside Claude Code or Hermes.

Verdict: watch. Score 22. A real competitor to track, not a fit to adopt while Claude Code remains YY's primary coding agent.

Source · 26.5k GitHub stars and 5k forks as of 2026-09-06, reached version 1.0 on August 7, 2026 per the release history, confirmed by direct fetch, prior review on 2026-07-16 caught it only at the closed beta stage.

Lians plugin

A small, evidence backed proof of done tool for Claude Code, Codex, and Cursor that runs real checks, binds the results to the current Git state, and reports what is actually ready for human review, available as a free local MCP server or a 10 dollar a month managed personal tier.

What it does for you: Matches the kind of check that render verify and Axion's own completion habits already push agents to do by hand, evidence tied to real Git state rather than a self reported claim, so it is worth a small trial on one Axion project to see if it catches anything the manual habit misses.

In practice: Small and early, but the honesty of its own published benchmarks, stating plainly that its cross agent continuity test is bounded beta evidence and not a competitor leaderboard, is a good sign for a tool about proving things are actually done.

For: Claude (Axion). Ships as a local MCP server built specifically for Claude Code, Codex, and Cursor sessions, so it plugs directly into Axion's existing Claude based workflow rather than a separate runtime.

Verdict: test first. Score 22. Directly matches YY's own verification before completion habit, low cost free local tier makes a trial cheap even though the project itself is tiny.

Source · 10 GitHub stars and 5 forks as of 2026-09-06, published a bounded beta result of 10 out of 10 expected facts recovered with 0 stale facts reported as current in its own cross agent continuity fixture, confirmed by direct fetch.

TEN-framework repo

An established open source framework for real time, multimodal conversational voice AI agents, Apache 2.0 licensed with some additional restrictions, that passed a public 6,000 GitHub star milestone earlier in 2026.

What it does for you: No live voice product to attach it to today, so this is background context rather than a build, useful mainly as a comparison point if MCL ever wants a voice agent beyond the existing IG story and hook writer content pipeline.

In practice: A real, established project rather than a new entrant, first picked up by this scout today.

For: Both runtimes. A multi language voice agent framework with C, C++, Go, Python, and JS or TS support, not tied to any single agent runtime.

Verdict: watch. Score 21. No current voice product need, tracked as an established reference option.

Source · Passed a public 6,000 GitHub star milestone as announced by the project earlier in 2026, Apache 2.0 licensed with additional restrictions noted in its own license file, confirmed by direct fetch on 2026-09-06.

qwen-audio-agent repo

A realtime voice runtime that keeps an agent talking, working, and present, Apache 2.0 licensed, shipping frequent dated releases through August 2026 including embeddable gateway and realtime provider extensions, Agent Skills support, and a version 2.0.0 now in development.

What it does for you: Low direct fit today since MCL and Axion work in English and the project's docs and community lean Chinese first, but the shipping cadence, six releases in about a month, is a useful signal of how fast realtime voice agent tooling is moving.

In practice: Actively developed and shipping real features weekly, but built for a different language and product context than YY's current work.

For: Standalone tool. A realtime voice product built around Qwen's own models and desktop app, not something Axion or Hermes would embed directly.

Verdict: watch. Score 19. Low fit for MCL's English first, non voice product today, tracked for its shipping pace rather than for adoption.

Source · 2.4k GitHub stars and 208 forks as of 2026-09-06, Apache 2.0 license, version 1.11.0 released 2026-08-20 and version 2.0.0 in development as of 2026-08-27, confirmed by direct fetch.

opentrace plugin

A small, single author observability engine for AI coding agents, self hosted on roughly a 4 dollar a month VM, storing structured logs, request lifecycle data, and error traces in a custom columnar store with no dashboard, designed so the AI assistant itself queries production data through MCP instead of a human copying and pasting it in.

What it does for you: An interesting shape for Axion's own bridge and heartbeat monitoring, letting an agent query its own production logs directly through MCP rather than YY manually tailing runs.log, but with only 25 stars and one author it needs a real trial before any trust.

In practice: A clever, narrow idea, no dashboards, the agent is the interface, executed by one person rather than a team.

For: Both runtimes. A self hosted, MCP native observability server that any MCP capable client can query, not tied to Claude specifically.

Verdict: watch. Score 18. Interesting shape for Axion's own observability but too new and single author to trust without a trial first.

Source · 25 GitHub stars as of 2026-09-06, self hosted on roughly a 4 dollar a month VM per its own README, single binary with SQLite storage, confirmed by direct fetch.

omlx (jundot) repo

An LLM inference server for Apple Silicon Macs with continuous batching and SSD backed KV caching, managed from the macOS menu bar, Apache 2.0 licensed.

What it does for you: Nothing directly, YY's box is Windows 11 per the standing environment rules, so this cannot run here at all, it is included only because it surfaced strongly in this week's local inference sweep and is worth knowing about if that ever changes.

In practice: A genuinely popular, well built project, the mismatch is entirely about YY's own hardware, not the project's quality.

For: Standalone tool. A macOS only, Apple Silicon specific inference server, and YY's box is Windows 11, so it does not run in Axion's actual environment at all.

Verdict: watch. Score 17. High quality project but a hard platform mismatch with YY's Windows box rules out any near term use.

Source · 21.5k GitHub stars and 1.8k forks as of 2026-09-06, Apache 2.0 license, confirmed by direct fetch.

cordon (marras0914) plugin

A security gateway for MCP tool calls that sits between an LLM client and MCP servers, adding sequence aware policies, an audit log of every call, and human approval through a terminal prompt or Slack before a flagged tool call runs.

What it does for you: The human approval before a dangerous tool call pattern is the same shape as Axion's own GHL write gate hook, approve once, approve twice, so it is conceptually close to something YY already does by hand, but the project itself has essentially no adoption yet.

In practice: Well documented for its size, with a genuinely useful sounding replay feature for late approvals, but one star and one fork is a sign nobody else has tried it yet.

For: Claude (Axion). Positioned as a gateway between an LLM client such as Claude Code and the MCP servers it calls, so it is specific to that MCP client pattern.

Verdict: skip. Score 15. Concept matches YY's own approval gate pattern, but one star, one fork, and a single author is too little evidence to trust with real tool calls today.

Source · 1 GitHub star and 1 fork as of 2026-09-06, confirmed by direct fetch, documents a Slack approval flow with OAuth and a replay command for late approvals.

awesome-agent-plugins (ZeroPointRepo) repo

A curated directory listing more than 40 plugins verified against the new Agent Plugins open standard covered at rank 1 today.

What it does for you: A useful bookmark for finding real Agent Plugins as the ecosystem grows, but it is a list, not a tool, so it has nothing to build against on its own.

In practice: A thin companion project to the real news this week, useful as a pointer rather than as a thing in itself.

For: Both runtimes. A markdown directory of links, not runtime specific at all.

Verdict: skip. Score 14. A directory, not a tool, useful only as a discovery aid for rank 1's ecosystem.

Source · 5 GitHub stars and 5 forks as of 2026-09-06, confirmed by direct fetch, lists itself as tracking more than 40 verified plugins for the Agent Plugins standard.

vigil (VectorInstitute) repo

A kernel level observability and enforcement tool for AI agents from the Vector Institute, a Canadian AI research lab, using BPF LSM hooks to watch and enforce policy on agent behavior at the operating system level.

What it does for you: Conceptually the deepest kind of sandboxing found this week, enforcement below the process itself rather than around it, but it needs a Linux kernel 5.7 or newer with BPF LSM enabled and root access, none of which matches YY's Windows 11 box.

In practice: A credible research lab behind it, but this is very early, three stars, and simply cannot run in Axion's actual environment today.

For: Standalone tool. Requires a Linux kernel with BPF LSM enabled and root privileges to load eBPF programs, which does not match YY's Windows 11 box at all.

Verdict: skip. Score 13. Credible origin but a hard platform mismatch with YY's Windows box plus very early stage traction rules out any near term use.

Source · 3 GitHub stars and 0 forks as of 2026-09-06, requires Linux kernel 5.7 plus with CONFIG_BPF_LSM enabled and root privileges per its own README, confirmed by direct fetch.