Daily AI Implementation Scout Council

2026-08-20. Top pick: #1 mattpocock/skills. Each item is graded on 7 axes; copy a build command to act on it.

Today's ranked top 20

#1
build nowBoth runtimes33 / 35

mattpocock/skills skill

Engineering skills from Matt Pocock's personal .agents directory, published as a public repo. Includes ask-matt (skill router), grill-with-docs (domain model builder that updates CONTEXT.md and ADRs inline), triage (issue state machine), improve-codebase-architecture (visual HTML report of deepening opportunities), to-spec (formalises the current conversation into a spec), to-tickets (decomposes any plan into issue-tracker tickets), and setup-matt-pocock-skills (one-time repo config). MIT license, Shell.

What it does for you: AXION engineering decisions, spec writing, and codebase architecture reviews are currently ad hoc. These skills add a structured workflow layer: to-spec captures decisions already in the chat into a formal spec file; to-tickets decomposes that spec into trackable work items; improve-codebase-architecture scans any AXION component and produces a visual report of where depth can be added. Matt Pocock uses these skills in his own daily engineering work, making them tested against real codebases rather than generic templates.

In practice: 223,647 GitHub stars, 19,244 forks, 1,214 stars on 2026-08-20, the number-one trending repo across all languages today. Matt Pocock is the author of Total TypeScript and a practising TypeScript educator. The description says these skills come straight from his .agents directory, which is the clearest possible provenance signal for a skill collection.

For: Both runtimes. SKILL.md files with no build step. Compatible with Claude Code, Hermes, and any skill-runner that reads the Agent Skills format.

Security4
Quality5
Auditability5
Useful to you5
Useful to community5
Buildable now5
Hermes4

Verdict: build now. 33 total. Clone and copy individual SKILL.md files into AXION/Skills today with no infrastructure requirement. The to-spec and to-tickets skills fill a gap that has come up repeatedly in AXION planning sessions. setup-matt-pocock-skills must run once per repo first.

Build #1 mattpocock/skills: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/mattpocock/skills. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 223,647 GitHub stars, 19,244 forks, 1,214 stars on 2026-08-20 (number-one trending all-languages). MIT license. Shell. Published from author's personal .agents directory.

#2
newtest firstBoth runtimes29 / 35

mukul975/Anthropic-Cybersecurity-Skills skill

817 structured cybersecurity skills for AI agents, mapped to 6 frameworks: MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3 Fight Fraud. Covers 29 security domains including threat detection, vulnerability management, red-team C2, phishing simulation, incident response, and compliance. Apache 2.0. Community project; not affiliated with Anthropic despite the name.

What it does for you: AXION has no dedicated security skills today. This library adds 817 structured audit and defence skills that Claude Code can invoke when reviewing AI-generated code, checking AXION scripts for vulnerabilities, or auditing MCL platform changes against known attack patterns. Start with the NIST CSF and D3FEND domains for pure-defence use: those skills cover asset inventory, access control review, and log analysis with no offensive techniques involved.

In practice: 29,781 GitHub stars, 3,530 forks, 767 stars on 2026-08-20, the number-one trending Python repo today. Apache 2.0. Security.md and Code of Conduct included. The MITRE mapping is unusually rigorous for a community skills repo.

For: Both runtimes. agentskills.io standard SKILL.md format. Works with Claude Code, GitHub Copilot, Codex CLI, Cursor, Gemini CLI, and 20+ platforms. Hermes can load SKILL.md files directly.

Security3
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now5
Hermes4

Verdict: test first. 29 total. The defence and CSF domains are immediately applicable to AXION security reviews. Test on a non-production AXION review first, picking only defence-domain skills. Do not install offensive-technique skills without explicit written authorization for the target system.

Build #2 mukul975/Anthropic-Cybersecurity-Skills: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/mukul975/Anthropic-Cybersecurity-Skills. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 29,781 GitHub stars, 3,530 forks, 767 stars on 2026-08-20 (number-one trending Python). 817 skills across 29 domains. 6 framework mappings documented in README. Apache 2.0.

#3
build nowBoth runtimes32 / 35

microsoft/SkillOpt repo

Text-space optimizer from Microsoft Research that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits and validation-gated updates. Produces deployable best_skill.md artifacts. Ships SkillOpt-Sleep in v0.2.0: a nightly self-evolution engine that harvests real agent session data, mines improvement opportunities, replays and consolidates them behind a held-out validation gate. Claude Code integration shell is in the repo.

What it does for you: AXION updates skills manually. SkillOpt replaces that loop with a research-backed automated pipeline: SkillOpt-Sleep reads real Claude Code session trajectories, identifies skill weaknesses, proposes edits, validates them against held-out tests, and writes an improved best_skill.md. This is the first skill optimizer with a published paper, a PyPI release, and a Claude Code integration shell in one package.

In practice: 16,100 GitHub stars, 1,500 forks as of 2026-08-19. Microsoft Research blog feature 2026-07-24 with coverage from VentureBeat and The Decoder. Re-listed from 2026-08-19 as a build-now carry: not yet adopted in AXION.

For: Both runtimes. Python package (pip install skillopt). Ships integration shells for Claude Code, Codex, Copilot, and Devin. SKILL.md artifacts are format-agnostic and run in Hermes or any skill-compatible harness.

Security4
Quality5
Auditability5
Useful to you5
Useful to community5
Buildable now4
Hermes4

Verdict: build now. 32 total. Re-listed from 2026-08-19 because no material change is required to justify keeping a build-now in the action queue for a second consecutive day. The install path (pip install skillopt plus the Claude Code shell in the repo) is same-day safe.

Build #3 microsoft/SkillOpt: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/microsoft/SkillOpt. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 16,100 stars, 1,500 forks as of 2026-08-19. v0.2.0 on PyPI released 2026-07-02. Microsoft Research blog post 2026-07-24. MIT license. Last listed 2026-08-19 with verdict build now.

#4
build nowBoth runtimes32 / 35

alirezarezvani/claude-skills skill

Community collection of 345 Claude Code and agent skills across 10 domains: engineering core (52 skills), engineering powerful (84 skills including agent-harness, zero-hallucination-coder, and skillopt-sleep), product, marketing, compliance, C-level advisory, research, business operations, commercial and finance, and daily productivity. Ships 30+ prebuilt agents and 70+ custom commands.

What it does for you: YY can clone this repo and cherry-pick from 345 battle-tested skills rather than writing from scratch. The agent-harness skill adds a goal-plan-execute-verify-close loop over any domain. The engineering/skillopt-sleep module ties this collection to the SkillOpt automated nightly improvement loop at rank 3. Updated 2026-08-18, two days before today.

In practice: 24,600 GitHub stars, 3,500 forks as of 2026-08-19. Updated 2026-08-18. The breadth at 345 skills across a domain table in the README signals systematic build, not ad hoc collection. Re-listed from 2026-08-19 as a build-now carry: not yet adopted in AXION.

For: Both runtimes. Markdown skill files with no build step; compatible with Claude Code, Codex, Hermes, and any skill-runner that reads SKILL.md or CLAUDE.md format.

Security4
Quality4
Auditability4
Useful to you5
Useful to community5
Buildable now5
Hermes5

Verdict: build now. 32 total. Re-listed from 2026-08-19 because no material change is required to justify keeping a build-now in the action queue for a second consecutive day. Clone and copy individual skills today.

Build #4 alirezarezvani/claude-skills: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/alirezarezvani/claude-skills. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 24,600 GitHub stars, 3,500 forks as of 2026-08-19. Updated 2026-08-18. 345 skills across 10 domains. Last listed 2026-08-19 with verdict build now.

#5
test firstBoth runtimes31 / 35

pydantic/pydantic-ai repo

Python AI agent framework from the Pydantic team. Typed end-to-end across agents, realtime voice, image generation, and embeddings. Every major model provider supported, including Anthropic. Dependency injection for tools. Pydantic model validation on all LLM outputs. Supports structured streaming and multi-turn conversations with typed state.

What it does for you: AXION agents today produce untyped output requiring ad hoc JSON parsing and retry logic. Pydantic AI replaces that pattern with full typing from the LLM response through to Python objects, eliminating custom parsing code in every AXION agent. The Anthropic provider ships natively; no adapter or wrapper needed. Realtime voice is a path to future MCL voice coaching tools.

In practice: 19,400 GitHub stars, 2,600 forks as of 2026-08-20. MIT license. Built by the Pydantic team, whose validation library is a Python ecosystem standard. Last seen in registry 2026-08-17 with verdict watch; upgraded to test first today based on the rising adoption signal and the gap in AXION's structured output handling.

For: Both runtimes. Python package; provider-agnostic typed agent framework. Native Anthropic provider ships in the package. Hermes can use it directly for typed response validation from any LLM call.

Security4
Quality5
Auditability5
Useful to you4
Useful to community4
Buildable now4
Hermes5

Verdict: test first. 31 total. The Anthropic provider is native, making this a same-session test. Try it on one AXION agent that currently does manual JSON parsing and compare the output quality before committing to a wider adoption.

Build #5 pydantic/pydantic-ai: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/pydantic/pydantic-ai. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 19,400 GitHub stars, 2,600 forks as of 2026-08-20. MIT license. Anthropic provider documented at ai.pydantic.dev/api/models/anthropic/. pip install pydantic-ai. Last seen in registry 2026-08-17.

#6
newtest firstBoth runtimes28 / 35

comet-ml/opik repo

Open-source LLM observability, evaluation, and monitoring platform from CometML. Captures multi-step agent traces: LLM calls, tool executions, retrieval steps, and agent activity in a searchable trace tree. Includes LLM-as-a-judge evaluation metrics, production dashboards, online evaluation rules, guardrails, and a PyTest integration for CI/CD evaluation on every commit. Ships an Agent Optimizer SDK for prompt and agent improvement.

What it does for you: AXION runs daily agent routines with no visibility into token usage, latency, prompt quality, or eval scores between runs. Opik adds an always-on trace layer that captures every AXION agent run in a searchable dashboard. YY can compare scout run quality across days, debug slow or failing nightly routines, and run automated evals on the scout newsletter output before it deploys. The PyTest integration means future AXION eval tests can gate the nightly pipeline.

In practice: 21,500 GitHub stars, 1,700 forks as of 2026-08-20. MIT license. CometML is a known ML experiment tracking company. First appearance in this registry. The CI/CD eval integration and the LLM-as-a-judge metrics are the two features most lacking in AXION's current pipeline.

For: Both runtimes. Python SDK plus self-hosted Docker or Opik Cloud. AXION or Hermes agents call it via a decorator or SDK wrapper. Provider-agnostic: works with any LLM.

Security4
Quality4
Auditability4
Useful to you4
Useful to community4
Buildable now4
Hermes4

Verdict: test first. 28 total. Self-hosted Docker install is a one-command start. Test by wrapping one AXION scout run with Opik tracing and reviewing the trace output before committing to full integration.

Build #6 comet-ml/opik: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/comet-ml/opik. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 21,500 GitHub stars, 1,700 forks as of 2026-08-20. MIT license. pip install opik. Self-hosted via Docker or Opik Cloud. First seen in this registry 2026-08-20.

#7
test firstClaude (Axion)30 / 35

diet103/claude-code-infrastructure-showcase repo

Reference implementation of Claude Code infrastructure patterns: a skill-activation-prompt hook that fires on every user prompt and checks a skill-rules.json file for trigger patterns, suggesting relevant skills automatically. Ships a hooks README, agents README, skills README, and a Claude integration guide written for AI-assisted setup.

What it does for you: AXION already has hooks in settings.json but no systematic trigger-based skill activation. Copying the two essential hooks and one skill-rules.json from this repo introduces declarative, auditable auto-activation. This is particularly relevant now that AXION has a large number of skills from ranks 1, 2, 3, and 4 above: a skill-rules.json makes routing to the right skill automatic instead of manual.

In practice: 10,000 GitHub stars as of 2026-08-19. Updated 2026-08-19. TypeScript, MIT. Re-listed from 2026-08-19 as a test-first carry: not yet adopted in AXION.

For: Claude (Axion). Claude Code-specific: the skill-activation-prompt hook and skill-rules.json pattern are designed for Claude Code's hook and skill runtime. Hermes does not run the same hook lifecycle.

Security4
Quality4
Auditability4
Useful to you5
Useful to community4
Buildable now5
Hermes4

Verdict: test first. 30 total. Re-listed from 2026-08-19 because the skill library at ranks 1, 2, 3, and 4 makes the auto-activation pattern more urgent than yesterday. Test on an isolated project first to confirm the hook does not conflict with the existing build-gate.js hook already in AXION settings.json.

Build #7 diet103/claude-code-infrastructure-showcase: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/diet103/claude-code-infrastructure-showcase. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 10,000 GitHub stars as of 2026-08-19. Updated 2026-08-19. TypeScript, MIT license. Ships CLAUDE_INTEGRATION_GUIDE.md plus three README files covering skills, hooks, and agents. Last listed 2026-08-19.

#8
test firstBoth runtimes29 / 35

akitaonrails/ai-memory repo

Rust CLI for long-term memory across agent coding sessions and cross-vendor handoff. Stores structured task context, failed approaches, open questions, and architecture notes in a vendor-agnostic format. When you switch from Claude Code to another CLI in the same directory, the next agent reads the memory and continues without re-explaining the project. Ships a web UI for memory inspection and proactive query routing via Agent Skills.

What it does for you: AXION sessions switch between Claude Code and Hermes; context is currently lost unless YY writes handoffs manually. ai-memory automates that handoff: after any Claude Code session the memory updates, and the next session (Claude or Hermes) picks it up automatically. The Agent Skills integration means routing guidance travels with the memory file.

In practice: 2,673 GitHub stars as of 2026-08-19. Rust 1.95+, MIT license. Re-listed from 2026-08-19 as a test-first carry: not yet adopted in AXION.

For: Both runtimes. Rust CLI binary that writes a vendor-agnostic memory block and Agent Skills files into any project directory. Works with Claude Code, Codex, Gemini CLI, and any agent CLI. Hermes can read the resulting memory files at session startup.

Security4
Quality4
Auditability4
Useful to you5
Useful to community5
Buildable now4
Hermes3

Verdict: test first. 29 total. Re-listed from 2026-08-19. Directly solves AXION's cross-session memory loss. Test on one non-critical project first to confirm the memory format does not expose sensitive AXION architecture in a shared directory.

Build #8 akitaonrails/ai-memory: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/akitaonrails/ai-memory. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 2,673 stars, 231 forks, trending Rust as of 2026-08-19. Rust 1.95+, MIT license. Release badge visible in README. Last listed 2026-08-19.

#9
test firstClaude (Axion)28 / 35

context7 (claude-plugins-official external plugin) plugin

MCP server that injects up-to-date library documentation into Claude Code context. When code references a library, context7 fetches current docs for that library version and adds them as context, preventing hallucinated API calls based on outdated training data. Listed in the Anthropic official external plugins directory.

What it does for you: AXION skills and agents frequently call Python library APIs that change between versions. context7 eliminates the class of errors where Claude generates code for an outdated API signature. Anthropic has reviewed this integration for listing in the official directory. Not currently in YY's enabledPlugins.

In practice: Listed in the Anthropic-managed external_plugins directory alongside GitHub, GitLab, Discord, Asana, and Firebase. Widely referenced in the Claude Code community as a standard quality-of-life install. No new entries in the official plugins directory since yesterday.

For: Claude (Axion). MCP server installed as a Claude Code external plugin via anthropics/claude-plugins-official. Not applicable to Hermes directly.

Security4
Quality4
Auditability4
Useful to you4
Useful to community4
Buildable now5
Hermes3

Verdict: test first. 28 total. Re-listed from 2026-08-19; still not in YY's enabledPlugins. Directly installable via the official plugin directory. The documentation accuracy improvement is measurable immediately.

Install: add context7 to enabledPlugins in ~/.claude/settings.json following the external_plugins/context7 README in anthropics/claude-plugins-official. Not currently in YY's enabledPlugins. Review before enabling.

Source · Listed at github.com/anthropics/claude-plugins-official/tree/main/external_plugins/context7 as of 2026-08-20. Anthropic-reviewed. 6 external plugins in the directory as of today: asana, context7, discord, fakechat, firebase, github, gitlab. Last listed 2026-08-19.

#10
test firstClaude (Axion)28 / 35

github (claude-plugins-official external plugin) plugin

Official GitHub MCP server for Claude Code. Provides direct GitHub API access from within a Claude Code session: read repos, files, issues, and PRs; create branches; commit files; open PRs. Listed in the Anthropic official external plugins directory.

What it does for you: AXION runs git operations manually via Bash tools today. The GitHub plugin lets Claude Code read issues, create branches, and open PRs directly, reducing tool-call overhead for the AXION docs repo and future tracked work. Not currently enabled in YY's settings.

In practice: Listed in the Anthropic-managed external_plugins directory. GitHub integration is the most-requested MCP server type across the community. No new entries in the official plugins directory since yesterday.

For: Claude (Axion). MCP server installed as a Claude Code external plugin via anthropics/claude-plugins-official. Not applicable to Hermes directly.

Security3
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now5
Hermes3

Verdict: test first. 28 total. Re-listed from 2026-08-19; still not in YY's enabledPlugins. Test on a non-critical repo first to confirm write behaviour before enabling on the main AXION docs repo.

Install: add the github external plugin to enabledPlugins in ~/.claude/settings.json following the external_plugins/github README in anthropics/claude-plugins-official. Not currently in YY's enabledPlugins. Review before enabling.

Source · Listed at github.com/anthropics/claude-plugins-official/tree/main/external_plugins/github as of 2026-08-20. Anthropic-reviewed external plugin. Last listed 2026-08-19.

#11
newwatchClaude (Axion)26 / 35

MadsLorentzen/ai-job-search repo

Claude Code-native job application framework: evaluate job postings, tailor CVs, write cover letters, prepare for interviews. Runs locally in Claude Code. Fork it and adapt it. Python, MIT license.

What it does for you: Not core AXION infrastructure. This repo is a textbook reference for a production-quality Claude Code multi-step pipeline: it chains evaluation rubrics, PDF generation via Playwright, structured storage, and domain-specific commands into a coherent Claude Code skill set. Study the architecture before designing the next AXION pipeline feature that chains rubric scoring, document output, and a tracker.

In practice: 32,246 GitHub stars, 11,308 forks, 166 stars on 2026-08-20, trending today. Python, MIT. The 11k forks signal it is being actively forked and adapted, not just starred.

For: Claude (Axion). Claude Code-native: the agent patterns, slash commands, and PDF pipeline are built for the Claude Code CLI runtime. Not directly usable in Hermes.

Security4
Quality4
Auditability4
Useful to you3
Useful to community5
Buildable now4
Hermes2

Verdict: watch. 26 total. Not AXION infrastructure; useful_yy score of 3 reflects pattern-study value. Watch for 30 days to see if any AXION pipeline feature needs a similar chain before building it.

Build #11 MadsLorentzen/ai-job-search: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/MadsLorentzen/ai-job-search. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 32,246 GitHub stars, 11,308 forks, 166 stars on 2026-08-20, trending today. Python, MIT license. First appearance in this registry.

#12
newwatchClaude (Axion)26 / 35

santifer/career-ops repo

The most fully featured open-source AI job search framework, with an A-G evaluation rubric, portal scanner for 100+ companies including Anthropic, OpenAI, ElevenLabs, Retool, and n8n; cover letter and CV PDF generation via Playwright; an interview story bank that accumulates STAR stories across evaluations; salary negotiation scripts; and funded-company discovery via the company:funded command. CLI-agnostic. The first reference implementation of the CareerOps Manifesto.

What it does for you: Not directly AXION relevant. career-ops is the largest Claude Code multi-skill project trending today at 65,714 stars. The portal scanner and evaluation pipeline are reference architecture for any AXION pipeline that needs to scan a list of URLs, score each one against a rubric, and produce a structured PDF output. The CLI-agnostic CLAUDE.md design is a portable pattern for MCL outreach, lead research, or venue sourcing tools.

In practice: 65,714 GitHub stars, 12,735 forks, 193 stars on 2026-08-20, number-one trending JavaScript repo today. Documentation at career-ops.org. The A-G rubric and Block G legitimacy check are unusually rigorous for a community project.

For: Claude (Axion). CLI-agnostic: runs on Claude Code, Codex, OpenCode, Antigravity, or any OpenAI-compatible local model via CLAUDE.md skill format. Hermes could run it as a custom command if the CLAUDE.md is adapted.

Security4
Quality5
Auditability4
Useful to you2
Useful to community5
Buildable now4
Hermes2

Verdict: watch. 26 total. Not AXION infrastructure; useful_yy score of 2 reflects pattern-study value only. The portal scanner architecture is worth studying for MCL research pipeline design.

Build #12 santifer/career-ops: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/santifer/career-ops. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 65,714 GitHub stars, 12,735 forks, 193 stars on 2026-08-20. JavaScript, MIT. Documentation at career-ops.org. First appearance in this registry.

#13
watchBoth runtimes26 / 35

agno-agi/agno repo

Agent platform SDK with 100+ pre-built tool integrations, cron-based scheduling, Telegram and WhatsApp interfaces, human approval workflow, JWT-based RBAC, multi-tenant isolation, and OpenTelemetry tracing. Designed to build and manage production agent platforms. 50+ production API endpoints exposed via SSE and websockets.

What it does for you: AXION uses Telegram for YY notifications, runs cron schedules via the bridge, and needs human approval gates for destructive operations. Agno ships all three as native primitives in a single SDK. The scheduling feature could replace the current Telegram bridge cron setup with a framework-managed scheduler, removing a maintenance surface.

In practice: 41,800 GitHub stars, 5,800 forks as of 2026-08-20. MIT license. Last seen in registry 2026-08-13 with verdict watch.

For: Both runtimes. Python SDK with a production API, 50+ endpoints, SSE and websocket support. AXION agents call it from Python; Hermes can submit work to an Agno-managed agent via HTTP.

Security3
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now3
Hermes3

Verdict: watch. 26 total. Re-listed after 7 days. Adopting Agno as the AXION runtime is a major architectural commitment that would replace the bridge and skill runner. buildable_now scored 3 because adoption requires code migration, not just installation.

Build #13 agno-agi/agno: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/agno-agi/agno. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 41,800 GitHub stars, 5,800 forks as of 2026-08-20. MIT license. pip install agno. Production API documented at docs.agno.com. Last seen in registry 2026-08-13.

#14
watchBoth runtimes27 / 35

langchain-ai/langgraph repo

Low-level graph-based orchestration framework for building stateful, long-running agents. Checkpointed state machine layer that survives failures and restarts. Production users include Klarna, Replit, and Elastic. The Deep Agents extension adds higher-level planning, subagent delegation, and file-system context for complex multi-step tasks. pip install -U langgraph.

What it does for you: AXION orchestrates multi-step workflows manually today, losing state on failure. LangGraph provides a checkpointed state machine layer that matches the same durability need as Temporal but with lighter deployment footprint for short-to-medium workflows. The Deep Agents extension handles the plan-execute-verify cycle that AXION currently implements by hand.

In practice: 40,000 GitHub stars, 6,700 forks as of 2026-08-20. MIT license. The production case studies at Klarna, Replit, and Elastic are the strongest validation available in the agent orchestration space. Last seen in registry 2026-08-15 with verdict test first.

For: Both runtimes. Python package; graph state management is provider-agnostic. Hermes can use LangGraph workflows directly. The Deep Agents extension built on top is also Python.

Security3
Quality5
Auditability4
Useful to you4
Useful to community5
Buildable now3
Hermes3

Verdict: watch. 27 total. Solid scores but buildable_now is 3 because adoption requires migrating AXION's existing orchestration code to the graph model. Held at watch until AXION has a concrete workflow migration target; upgrade to test first when that target is identified.

Build #14 langchain-ai/langgraph: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/langchain-ai/langgraph. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 40,000 GitHub stars, 6,700 forks as of 2026-08-20. MIT license. pip install -U langgraph. Production case studies for Klarna, Replit, Elastic at the LangGraph site. Last seen in registry 2026-08-15.

#15
newwatchBoth runtimes22 / 35

NVIDIA-NeMo/Switchyard repo

Rust proxy for LLM traffic from NVIDIA NeMo. Translates between OpenAI Chat, Anthropic Messages, and OpenAI Responses formats. Enables Claude Code or Codex to be routed to local models (vLLM, NVIDIA NIM, Ollama) without changing the agent's calling code. Supports multi-backend routing, A/B benchmarking, Prometheus metrics, and custom routing algorithms. Updated within minutes of today's scouting.

What it does for you: AXION today calls Claude directly for every request. Switchyard would let YY route cost-heavy batch tasks (nightly brain ingest, bulk eval scoring) to a local model while keeping Claude for interactive and quality-sensitive sessions. The Anthropic Messages format is natively supported on the output side, so no calling-code change is needed in Claude Code.

In practice: 1,900 GitHub stars as of 2026-08-20, updated within minutes of scouting. Rust. NVIDIA NeMo organization. Apache 2.0. New to this registry. Modest star count but NVIDIA-backed provenance is a strong signal for production readiness.

For: Both runtimes. Rust proxy that sits between any agent runtime and LLM providers. AXION or Hermes agents point at Switchyard instead of the Anthropic API; Switchyard translates and routes.

Security3
Quality3
Auditability4
Useful to you3
Useful to community3
Buildable now3
Hermes3

Verdict: watch. 22 total. The Anthropic-to-local-model routing is directly useful for AXION cost management, but 1,900 stars is modest and production stability for streaming, tool use, and vision needs independent verification before routing AXION sessions through it.

Build #15 NVIDIA-NeMo/Switchyard: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/NVIDIA-NeMo/Switchyard. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 1,900 GitHub stars as of 2026-08-20, updated within minutes of scouting today. Rust. NVIDIA NeMo organization. Apache 2.0 license. First appearance in this registry.

#16
watchBoth runtimes27 / 35

huggingface/smolagents repo

Barebones Python library for agents that plan in code. CodeAgent generates executable Python plans that run locally; ToolCallingAgent uses standard tool-call JSON. LLM-agnostic: HuggingFace inference, Anthropic via LiteLLM, Together AI, OpenAI-compatible servers, and local models all supported. Agents can be pushed to and loaded from the HuggingFace Hub. pip install smolagents.

What it does for you: smolagents is the simplest possible agent framework: no heavy abstraction, agents plan and execute as Python code. For AXION scripts that are currently written as procedural Python, smolagents offers a path to agentic execution with minimal framework overhead. The Hub push/pull means AXION agents can be versioned and shared across machines, aligning with the Mac brain sync setup.

In practice: 28,900 GitHub stars, 2,900 forks as of 2026-08-20. MIT license. HuggingFace official repo. The Hub push/pull and multi-provider support are features no other barebones framework ships together.

For: Both runtimes. Python package, LLM-agnostic. Anthropic Claude is supported via the LiteLLMModel wrapper (LiteLLM is a separate dependency). Hermes can use smolagents workflows directly from Python.

Security4
Quality4
Auditability4
Useful to you3
Useful to community5
Buildable now4
Hermes3

Verdict: watch. 27 total. Claude requires the LiteLLMModel wrapper rather than a native provider, adding a dependency layer. Watch for a native Anthropic provider or until AXION has a concrete use case for a code-planning agent.

Build #16 huggingface/smolagents: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/huggingface/smolagents. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 28,900 GitHub stars, 2,900 forks as of 2026-08-20. MIT license. pip install smolagents. Native Anthropic path uses LiteLLM wrapper per the README.

#17
newwatchBoth runtimes24 / 35

lm-sys/RouteLLM repo

Framework for serving and evaluating LLM routers. Routes queries to a strong or weak model based on task difficulty. Trained routers classify whether a query needs a powerful model or can be served by a cheaper one. Drop-in replacement for the OpenAI client. Published paper at arxiv.org/abs/2406.18665 (LMSYS). Claims 85% cost reduction while maintaining 95% GPT-4 performance on MT Bench benchmarks.

What it does for you: AXION runs every query through Claude regardless of complexity. RouteLLM would classify low-complexity tasks (formatting, simple lookups, routine summaries in the brain ingest pipeline) to a cheaper model and route only complex reasoning to Claude, reducing daily API costs across AXION's agent runs.

In practice: 5,400 GitHub stars, 421 forks as of 2026-08-20. MIT license. From LM-Sys, the team behind LMSYS Chatbot Arena. Published research paper with benchmark comparisons across MT Bench.

For: Both runtimes. Python package; acts as a drop-in replacement for the OpenAI client or as a standalone server. AXION or Hermes agents call it via the same interface as the current API client.

Security3
Quality4
Auditability4
Useful to you3
Useful to community3
Buildable now4
Hermes3

Verdict: watch. 24 total. The cost reduction claim of 85% is compelling but the trained routers were benchmarked on GPT-4 vs. a weaker model; Claude-to-local-model routing needs separate validation. Watch until AXION has a concrete cost-reduction target to test against.

Build #17 lm-sys/RouteLLM: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/lm-sys/RouteLLM. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 5,400 GitHub stars, 421 forks as of 2026-08-20. MIT license. pip install routellm. Published paper at arxiv.org/abs/2406.18665. 85% cost reduction claim on MT Bench documented in paper and blog post.

#18
test firstClaude (Axion)29 / 35

modelcontextprotocol/servers plugin

Anthropic-maintained collection of reference and community MCP server implementations: filesystem, GitHub, Slack, Google Drive, Postgres, fetch, memory, and others. The canonical starting point for adding MCP tools to Claude Code. Re-listed for the second consecutive day; no new servers added today.

What it does for you: AXION uses Claude Code with MCP tools daily. The memory server ships a knowledge graph implementation useful for AXION agent memory. The fetch server is directly relevant for the scout newsletter pipeline. Both can be installed from this repo without a build step.

In practice: Continuously maintained by Anthropic and the MCP community. No new entries observed in today's plugin scout. Re-listed as a daily reminder that the memory and fetch servers have not yet been evaluated for AXION.

For: Claude (Axion). Official MCP server reference implementations; each server is installed independently into Claude Code or any MCP client via the mcpServers block in settings.json.

Security4
Quality5
Auditability5
Useful to you3
Useful to community5
Buildable now4
Hermes3

Verdict: test first. 29 total. Re-listed from 2026-08-19. The memory and fetch servers score buildable_now at 5 individually; evaluate those two specifically rather than the full collection.

Install: enable individual servers by adding them to the mcpServers block in ~/.claude/settings.json following each server's README. Review before enabling.

Source · Official Anthropic repository. MCP org has 24,000+ GitHub followers. Memory server and fetch server documented in the README with npx install commands. No new servers observed in today's plugin scout. Last listed 2026-08-19.

#19
watchBoth runtimes27 / 35

run-llama/llama_index repo

Data framework for connecting LLMs to external data sources via indexing, retrieval, and query engines. Over-the-box connectors for documents, databases, APIs, and web pages. 100+ data source integrations. Structured loaders, vector and keyword indexing, and query pipelines with Anthropic Claude support built in. pip install llama-index.

What it does for you: AXION's brain ingest is a custom pipeline today handling PDFs, web pages, and documents with ad hoc parsing. LlamaIndex would standardize that pipeline: structured loaders handle format detection, indexing provides search, and the query engine returns cited answers. The Anthropic provider means no adapter is needed.

In practice: 40,000+ GitHub stars. MIT license. Re-listed after 7+ weeks (last seen 2026-07-03 with verdict test first). The ecosystem around LlamaIndex has matured; the query pipeline and data loaders are now production-stable per the docs.

For: Both runtimes. Python package; provider-agnostic data framework. Anthropic Claude is supported directly. Hermes can use LlamaIndex loaders and query engines from Python.

Security3
Quality4
Auditability4
Useful to you3
Useful to community5
Buildable now4
Hermes4

Verdict: watch. 27 total. Re-listed after 7+ weeks to check whether AXION's document ingestion priority has moved. buildable_now is 4 but adoption requires migrating the existing brain ingest pipeline. If document ingestion has not been prioritised in the next 30 days, this moves to skip.

Build #19 run-llama/llama_index: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/run-llama/llama_index. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 40,000+ GitHub stars. MIT license. pip install llama-index. Anthropic provider documented at docs.llamaindex.ai. Last seen in registry 2026-07-03 with verdict test first.

#20
watchBoth runtimes27 / 35

openai/openai-agents-python repo

OpenAI's lightweight multi-agent workflow SDK for Python. Provides agents, handoffs, guardrails, and tracing primitives. Updated 2026-08-19. Last seen in registry 2026-08-19 as watch.

What it does for you: YY designs multi-agent workflows for AXION. The handoff and guardrail patterns here are directly comparable to AXION's current orchestrator-worker design. The tracing integration ships structured trace output that could inform AXION agent logging. Updated yesterday while this issue was being compiled.

In practice: 28,800 GitHub stars as of 2026-08-19. OpenAI official SDK. Updated 2026-08-19. Re-listed for a second consecutive day because the update recency makes it the most current multi-agent SDK reference available.

For: Both runtimes. Python package; handoff and guardrail patterns apply to any agent framework including Hermes regardless of the underlying LLM provider.

Security3
Quality5
Auditability4
Useful to you3
Useful to community5
Buildable now4
Hermes3

Verdict: watch. 27 total. Re-listed from 2026-08-19 because it updated yesterday and remains the most-watched multi-agent SDK in this universe. Useful as a pattern reference; the OpenAI provider coupling limits direct adoption in AXION's Claude-first stack.

Build #20 openai/openai-agents-python: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/openai/openai-agents-python. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · 28,800 GitHub stars as of 2026-08-19. Updated 2026-08-19. OpenAI official SDK. MIT license. Last seen in registry 2026-08-19.