Daily AI Implementation Scout Council

Today's ranked top 20

#1

test firstBoth runtimes30 / 35

agentmemory repo

Persistent memory for AI coding agents: auto-captures sessions via hooks, stores locally in SQLite with the iii-engine, and exposes 51 MCP tools for remember, recall, recap, handoff, forget, commit-context, commit-history, and session-history operations.

What it does for you: agentmemory drops 12 hooks directly into Claude Code that capture every session, decision, and artefact without any manual save step, which is exactly the gap that causes repeated cold-start errors in Axion agent runs. When a new session begins, the recall hook surfaces exactly what the previous agent left off, cutting context-window burn and the time spent re-orienting to prior work. The 51-tool MCP server means Hermes can read and write the same memory store as Claude Code, giving both runtimes one shared knowledge base.

In practice: Benchmark-first, zero external dependencies, opinionated in the best way. Feels like the missing persistence layer every Claude Code workflow has needed.

For: Both runtimes. Zero-dependency local-first library; 12 auto-capture hooks ship for Claude Code, 6 for Codex CLI, and a 51-tool MCP server covers every other client including Hermes.

Security4

Quality5

Auditability4

Useful to you5

Useful to community5

Buildable now5

Hermes2

Verdict: test first. Score 30, one point below the 32 build-now threshold stated in dissent. 95.2% R@5 on LongMemEval-S is the strongest published retrieval benchmark in this class. Wire one Claude Code hook, confirm recall quality on a real Axion session, then promote to build now.

Build #1 agentmemory: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/rohitg00/agentmemory. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 9,361 GitHub stars confirmed, 95.2% R@5 on LongMemEval-S vs mem0 at 68.5% per benchmark/COMPARISON.md, 12 Claude Code auto-capture hooks documented, 51 MCP tools listed in README, npm install -g agentmemory install path confirmed, zero external dependencies stated in architecture docs.

#2

newtest firstClaude (Axion)29 / 35

GoodMemory plugin

Local-first auditable memory layer for Claude Code and Codex: read-only MCP recall by default, opt-in governed writeback, with inspect, forget, and export controls built in from day one.

What it does for you: GoodMemory adds durable session recall to Claude Code without any external server or cloud dependency. Every memory entry is auditable: inspect it, correct it, export it, or delete it at any time. The read-only default means Claude cannot silently overwrite prior context, which is the failure mode that causes stale memory corruption in less governed tools. The writeback opt-in matches the discipline YY already applies to brain ingests, making this the most compatible memory layer with the existing Axion governance model.

In practice: Privacy-first and governance-first. Smaller community than agentmemory but the cleanest audit story in the memory category.

For: Claude (Axion). Managed Claude Code and Codex CLI setup via goodmemory setup; standalone MCP mode for any other client. All memory stays in a local SQLite file by default.

Security5

Quality4

Auditability4

Useful to you5

Useful to community3

Buildable now5

Hermes3

Verdict: test first. Score 29. Excellent security and buildability profile for a hackathon-origin project. Install alongside agentmemory on a test Axion project and compare recall precision before standardising on one.

Install: run npm install -g goodmemory@0.7.5 then goodmemory setup for Claude Code or goodmemory-mcp --standalone --user-id YOUR_USER_ID for other MCP clients. Not currently in enabledPlugins. Review before enabling. Copy

Source · v0.7.5 published on npm, npm install -g goodmemory@0.7.5 install path confirmed, hackathon submission opened 2026-07-13, SQLite default store with Postgres opt-in documented in README, Codex and Claude Code setup commands confirmed.

#3

test firstBoth runtimes29 / 35

Phoenix v16 (Arize) repo

Open-source AI observability and evaluation platform: traces LLM calls across every provider, runs sandboxed server-side code evaluators for composite scoring, and exposes a remote MCP server so coding agents can query their own trace data while debugging.

What it does for you: Phoenix v16 adds sandboxed code evaluators that run Python scoring functions in kernel-level isolation on the server, so you can define a scoring rubric for any Axion agent output without installing dependencies locally. The remote MCP server is the new multiplier: Claude Code can query trace data and evaluation results mid-session, which turns debugging a broken agent from a copy-paste context exercise into a live tool call. For MCL content workflows that need reproducible quality scores rather than vibe-checks, this is the missing measurement layer.

In practice: The most complete open-source eval stack available. v16 sharpens the agent-driven debugging story significantly over v15.

For: Both runtimes. OpenTelemetry-based traces work from any Python or TypeScript runtime; a remote MCP server and Claude Code skill let coding agents query traces, datasets, and experiments directly.

Security4

Quality5

Auditability5

Useful to you4

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Score 29. Excellent review quality and auditability. Wire Phoenix traces into one Axion agent and confirm the MCP query works from Claude Code before broader adoption.

Build #3 Phoenix v16 (Arize): use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/Arize-ai/phoenix. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v16.0.0 released 2026-05-21, sandboxed code evaluators confirmed in release notes at arize.com, remote MCP server documented at arize.com/phoenix, built on OpenInference and OpenTelemetry standards, Claude Code skill published.

#4

test firstBoth runtimes29 / 35

Promptfoo repo

AI red-teaming and evaluation framework: runs LLM-as-judge evals, automated adversarial probes, and multi-turn conversation tests across providers. MIT license. Acquired by OpenAI March 2026 but remains fully open source.

What it does for you: Promptfoo lets you define expected outputs for every Axion and Hermes agent prompt, then run a regression test suite every time a prompt or model changes. The red-team mode generates adversarial inputs automatically, which surfaces the edge cases that manual testing on happy paths consistently misses. The OpenAI acquisition brings corporate maintenance backing without changing the license, so there is no lock-in risk for a Claude-first stack.

In practice: Battle-tested and now institutionally backed. The most accessible eval-and-red-team tool for an operator who does not want to build evaluators from scratch.

For: Both runtimes. CLI-first, runs against any LLM provider including Anthropic; invoked via npx promptfoo from any Claude Code script or Hermes shell step.

Security4

Quality5

Auditability5

Useful to you4

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Score 29. Excellent audit trail and quality scores. Run Promptfoo against the scout newsletter agent prompt and one MCL content agent as a proof of concept before standardising.

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

Source · MIT license confirmed on GitHub, acquired by OpenAI March 2026 per Openlayer guide, 17,600 GitHub stars and 2,900 forks as of January 2026, npx promptfoo install path documented, Anthropic provider supported in configuration docs.

#5

newtest firstBoth runtimes28 / 35

NOOA (NVIDIA) repo

NVIDIA Object-Oriented Agents: an alpha research preview that models an agent as a single Python class. Methods are capabilities, fields are state, docstrings are prompts, and type annotations are enforced contracts. Requires no orchestration layer, prompt registry, or external validator.

What it does for you: NOOA eliminates the orchestration boilerplate that clutters most Axion agent files: no separate prompt files, no tool registries, no validator layers wired by hand. Every agent you write is a standard Python class, so it is readable, testable, and composable with ordinary Python tooling without a framework-specific mental model. NVIDIA's internal benchmarks report roughly half the token spend on SWE-bench tasks versus standard scaffolding, which is a meaningful cost reduction if the result holds on Claude-backed Axion agents.

In practice: Minimal and elegant. The object-oriented framing makes complex agent state feel native rather than bolted on via configuration.

For: Both runtimes. Pure Python library, pip install nooa; runs in any Claude Code script or Hermes agent environment with Python 3.12 or 3.13.

Security4

Quality5

Auditability4

Useful to you4

Useful to community3

Buildable now4

Hermes4

Verdict: test first. Score 28. Alpha research preview with an unstable API surface. Rebuild one small Axion agent in NOOA and measure token spend and error rate before promoting.

Build #5 NOOA (NVIDIA): use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/NVIDIA-NeMo/labs-OO-Agents. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v0.0.8 released 2026-07-30, pip install nooa confirmed, Apache 2.0 license, 1,300+ GitHub stars within days of release, SWE-bench token reduction claim in NVIDIA release blog, Python 3.12 to 3.13 requirement documented.

#6

test firstBoth runtimes28 / 35

Microsoft Agent Framework 1.0 repo

GA-stable unified SDK that merges AutoGen and Semantic Kernel into one supported framework. Ships ChatAgent and Workflow primitives, enterprise observability, and the Foundry managed runtime for deployment.

What it does for you: If any Axion or Hermes dependency touches AutoGen or Semantic Kernel, MAF 1.0 is the supported migration target: one SDK, stable APIs, and Microsoft's support commitment behind it. AutoGen entered maintenance-only mode at Build 2026, meaning bug-fixes only and no new capabilities, so staying on AutoGen creates growing technical debt without a clear upgrade path. The Workflow primitive adds persistent state and graph-based orchestration comparable to LangGraph, now backed by a tier-1 vendor.

In practice: Consolidation release rather than an innovation release. Not exciting but important if any part of the stack still touches the Microsoft agent ecosystem.

For: Both runtimes. Ships in Python and .NET under MIT license; runs against any OpenAI-compatible endpoint including Anthropic models via a litellm bridge.

Security5

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now4

Hermes4

Verdict: test first. Score 28. Strong security posture and Python plus .NET parity. Run MAF ChatAgent alongside the Claude Agent SDK on one shared task and benchmark latency and reliability before committing.

Build #6 Microsoft Agent Framework 1.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/microsoft/agent-framework. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · MAF 1.0 GA on 2026-04-02 confirmed at devblogs.microsoft.com, AutoGen maintenance-only mode announced at Build 2026, MIT license, Python and .NET packages on PyPI and NuGet, Agent Harness GA confirmed alongside 1.0.

#7

newtest firstStandalone tool27 / 35

Orca ADE repo

Agent Development Environment by Stably AI (YC W22): run a fleet of parallel coding agents in one desktop application with Git Worktree isolation per agent, integrated terminals, editor, GitHub and Linear task workflows, and browser previews side by side.

What it does for you: Orca gives YY a dedicated UI for the multi-agent pattern already used in Axion: fan out tasks to 5 to 8 parallel agents, each in its own worktree, review diffs side by side, and merge. It replaces the manual worktree setup and terminal-switching overhead that slows parallel agent sessions down to a drag-and-drop workflow. YC W22 backing, weekly release cadence, and the fastest star growth of any developer tool on GitHub in 2026 suggest it will be maintained through production use.

In practice: The most polished ADE available right now. Feels like Cursor built for agent fleets rather than individual files.

For: Standalone tool. Desktop GUI application for managing fleets of parallel coding agents; not a library to embed in Axion code or call from Hermes.

Security3

Quality4

Auditability4

Useful to you5

Useful to community5

Buildable now4

Hermes2

Verdict: test first. Score 27. Standalone desktop GUI limits programmatic integration with Axion scripts; hermes_usable is low. Download and run one parallel task session to evaluate whether it replaces manual worktree setup before committing.

Build #7 Orca ADE: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/stablyai/orca. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 53,000+ GitHub stars by late August 2026, first commit 2026-03-17, approximately 10,000 new stars per month sustained since launch, YC W22 confirmed on Stably AI site, desktop and mobile builds confirmed live.

#8

newtest firstBoth runtimes27 / 35

Google ADK 2.0 repo

Google Agent Development Kit 2.0: graph-native workflow engine with stateful recovery, A2A 1.x protocol support, human-in-the-loop resumable execution, Skill registries, and the ability to expose any ADK agent as an MCP server.

What it does for you: ADK 2.0's agent-as-MCP-server export is the capability worth watching: any agent you build with ADK can now be surfaced as an MCP tool in the Axion layer without rewriting. A2A 1.x support means ADK agents can delegate to each other across Google Cloud services, relevant if MCL workloads ever need GCP scale. The graph-native workflow with state-based recovery makes it a credible LangGraph alternative with Google Cloud SLAs behind it.

In practice: Serious infrastructure play from Google. The MCP export closes the interoperability gap that made ADK feel isolated from the Claude ecosystem.

For: Both runtimes. Python and Go libraries; any ADK agent can be exposed as an MCP server, making it callable from Claude Code or Hermes via the standard MCP client.

Security5

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Score 27. MCP server export is the differentiator worth testing. Wire one small ADK agent as an MCP tool into Claude Code and confirm the handoff is clean before committing to the framework.

Build #8 Google ADK 2.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/google/adk-python. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · ADK 2.0 released 2026-06-30 confirmed in Go ADK release notes, Go ADK on Google GenAI v1.65.0 and Go 1.26 confirmed, A2A 1.x support documented at adk.dev, agent-as-MCP-server announced at Google I/O 2026.

#9

newtest firstBoth runtimes28 / 35

Composio MCP plugin

Managed MCP gateway with 1,000 plus pre-built integrations: OAuth, API key, and token-refresh management handled server-side. Gives Claude Code and the Claude Agent SDK access to business tools without writing any auth code.

What it does for you: Composio removes the credential-wiring step for every external service Axion needs to reach. Instead of building a custom MCP server for Slack, GHL, Google Sheets, or ClickUp, you point Claude Code at Composio's MCP endpoint and get all integrations through a single auth flow. The tool-level audit trail logs every external call, which satisfies the auditability requirement for MCL business automations. The Claude Agent SDK guide is published and verified.

In practice: Integration middleware done right. Trades self-host control for speed and breadth at a scale that would take months to build manually.

For: Both runtimes. MCP Gateway endpoint callable from Claude Code or Hermes via any MCP client; no custom integration code needed to reach any of the 1,000-plus managed tools.

Security3

Quality4

Auditability4

Useful to you5

Useful to community4

Buildable now4

Hermes4

Verdict: test first. Score 27. Excellent breadth and auth management, but SaaS dependency and rate limits need evaluation before MCL business automations rely on it. Test one GHL or ClickUp integration before wiring production workflows.

Install: add Composio MCP endpoint to .claude/settings.json mcpServers block using your Composio API key from composio.dev. Not currently in enabledPlugins. Review before enabling. Copy

Source · 1,000 plus managed integrations confirmed on composio.dev, Claude Code integration guide live at composio.dev/toolkits/composio/framework/claude-code, Claude Agent SDK guide published, MCP Gateway endpoint documented with per-team secure endpoints.

#10

test firstBoth runtimes26 / 35

Cognee repo

Graph-native AI memory platform: ECL pipeline (Extract, Cognify, Load) transforms raw data into queryable knowledge graphs. Ships 14 retrieval modes from classic RAG to graph-traversal chain-of-thought reasoning. August 2026 update adds Amazon Neptune as a managed graph backend.

What it does for you: Cognee's August update adds Amazon Neptune so MCL workloads that need production-scale graph traversal can run on AWS without standing up a self-hosted Kuzu instance. The 14 retrieval modes give a direct upgrade path from the flat-vector RAG currently used in the scout pipeline to graph-aware reasoning that understands relationships between tools, skills, and campaigns stored in the brain. At 12,000 stars and 70 plus production deployments, this is not a research project.

In practice: The most architecturally ambitious open-source memory project in the space. The graph framing is right even if the local setup is heavier than simpler alternatives.

For: Both runtimes. Python library with embedded defaults (SQLite, LanceDB, Kuzu): no external infrastructure needed to start; callable from Hermes via HTTP or direct Python import.

Security4

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Score 26. Neptune integration and 70-plus production deployments confirm real-world viability. Test against YY Brain graph queries before committing to the heavier infrastructure.

Build #10 Cognee: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/topoteretes/cognee. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 12,000 plus GitHub stars, 80 plus contributors, $7.5M seed round confirmed at cognee.ai, Amazon Neptune integration added August 2026 per changelog, 70 plus production deployments self-reported, VLDB 2026 architecture citation.

#11

newtest firstClaude (Axion)26 / 35

GitKraken Claude Plugin plugin

Official GitKraken plugin for Claude Code: gives Claude access to cross-platform Git data via the GitKraken MCP server. Covers commits, branches, pull requests, and issues across GitHub, GitLab, Azure DevOps, Bitbucket, and Jira.

What it does for you: If YY uses GitKraken as the primary Git interface, this plugin makes Claude Code aware of the full project context without switching tools or copy-pasting branch names. Claude can query PR status, commit history, and issue state in the same session where it is writing code, which removes a context-switch that typically costs a minute or two per task. No CLI install required: the plugin connects directly to the GitKraken MCP server.

In practice: Polished first-party integration. Low setup cost for GitKraken users; zero value if GitKraken is not in the workflow.

For: Claude (Axion). Claude Code plugin that surfaces GitKraken MCP server data; no CLI install required, runs inside Claude Code.

Security5

Quality4

Auditability4

Useful to you3

Useful to community4

Buildable now5

Hermes1

Verdict: test first. Score 25. Strong security and buildability, limited if YY does not use GitKraken as primary Git tool. Install and verify PR query latency on one active Axion repo before standardising.

Install: run claude plugins install gitkraken in terminal or find GitKraken under Claude Desktop > Settings > Plugins > Marketplace. Not currently in enabledPlugins. Review before enabling. Copy

Source · Plugin announced 2026-08-18 at gitkraken.com/blog, live at claude.com/plugins/gitkraken, GitHub repo at gitkraken/claude-plugin confirmed, supports GitHub, GitLab, Azure DevOps, Bitbucket, and Jira per announcement.

#12

newwatchStandalone tool24 / 35

vLLM v0.28.0 repo

High-throughput memory-efficient inference engine for LLMs: v0.28.0 adds request priority routing via HTTP header, session ID plumbing, Rust frontend with gRPC multimodal image inference, RL lifecycle control, and Anthropic API 4xx error correction for client-caused errors.

What it does for you: vLLM v0.28.0 is the most significant single release in this project's history by commit count: 584 commits from 270 contributors, 76 of them first-time. The Anthropic API 4xx correction means any Axion agent routing through a vLLM proxy will now receive properly typed client errors instead of ambiguous server errors, making debugging dramatically faster. Request priority routing matters if Axion ever runs mixed priority workloads through a shared inference server.

In practice: A major release masquerading as a version bump. The contributor count alone signals broad adoption and healthy maintenance.

For: Standalone tool. Inference server run as a standalone service; Claude Code or Hermes calls it via HTTP but cannot embed it in-process.

Security4

Quality5

Auditability4

Useful to you2

Useful to community5

Buildable now3

Hermes1

Verdict: watch. Score 24. Strong community and review quality, but useful_yy is limited: YY does not currently run a local inference server, and the Anthropic API is the primary model backend. Watch for the day a custom model or cost ceiling makes self-hosted inference necessary.

Build #12 vLLM v0.28.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/vllm-project/vllm/releases/tag/v0.28.0. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v0.28.0 released 2026-08-24, 584 commits from 270 contributors (76 new) confirmed in release notes, request priority header feature documented, Anthropic API 4xx correction confirmed in changelog, Hugging Face Inference Endpoints default to vLLM confirmed.

#13

watchBoth runtimes24 / 35

OpenViking repo

Self-evolving context database for AI agents from ByteDance Volcengine: stores memory, resources, and skills as one virtual filesystem under the viking:// protocol with ls, tree, and find navigation. Content is processed into three tiers: L0 abstract, L1 overview, L2 details.

What it does for you: OpenViking unifies what Axion currently manages separately, namely memory via the brain, skills via the Skills folder, and RAG context via ad-hoc fetches, into one addressable filesystem an agent can navigate natively. The three-tier content model (abstract, overview, detail) means an agent loads only the depth it needs per query, which is a principled approach to context-window management. VLDB 2026 acceptance provides academic peer review of the architecture.

In practice: Architecturally bold. The filesystem metaphor for agent context is the right abstraction even if AGPLv3 creates commercial use constraints.

For: Both runtimes. Python library open under AGPLv3; runs as a local context database accessible from Claude Code scripts or Hermes via the viking:// protocol.

Security3

Quality4

Auditability3

Useful to you4

Useful to community5

Buildable now3

Hermes2

Verdict: watch. Score 24. Strong community traction and a well-reasoned architecture, but AGPLv3 license and complex setup lower the score. Study the architecture before committing to the integration overhead.

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

Source · 30,000 plus GitHub stars and 2,300 plus forks confirmed at trendshift.io, open-sourced January 2026, AGPLv3 license confirmed, VLDB 2026 paper acceptance confirmed, viking:// protocol documented in repo README.

#14

newwatchBoth runtimes26 / 35

sqlite-memory repo

Markdown-based AI agent memory with semantic search, hybrid retrieval combining keyword and vector search, and offline-first sync between agents. Backed by SQLite with no external dependencies.

What it does for you: sqlite-memory stores agent memories as plain Markdown files in a SQLite database, which means every memory entry is human-readable and diffable without a special tool. Hybrid retrieval combining keyword and semantic search is more accurate than either alone on the kinds of structured queries Axion agents make about past work. Offline-first means no cloud sync failure can corrupt the memory store mid-session.

In practice: The most transparent memory store in the category. Markdown-native storage is a meaningful advantage for an operator who wants to audit and edit memories manually.

For: Both runtimes. Language-agnostic MCP server and CLI; any Claude Code script or Hermes agent connects via the MCP protocol or direct SQLite file access.

Security5

Quality3

Auditability3

Useful to you4

Useful to community3

Buildable now4

Hermes4

Verdict: watch. Score 24. Good security and Hermes usability scores, but community traction and review quality are lower than agentmemory and GoodMemory. Worth evaluating if the Markdown-native storage model matters for manual auditing.

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

Source · Repository at sqliteai/sqlite-memory confirmed live, SQLite FTS5 hybrid retrieval documented at sqlite.ai/sqlite-memory, offline-first architecture stated in README, MCP server and CLI access paths both documented.

#15

newwatchClaude (Axion)21 / 35

claudish-to-english plugin

Claude Code plugin that appends a plain-English rewrite under every Claude response, generated by a local LLM via Ollama. Default model is gemma4:26b-mlx for Apple Silicon; also supports the Anthropic API and any OpenAI-compatible endpoint.

What it does for you: When sharing Claude Code output in MCL team contexts or reviewing long agent responses quickly, claudish-to-english surfaces the one-sentence version of what Claude just said without having to read the full response. Because it runs locally via Ollama and is display-only, nothing leaves the machine and the original transcript is unchanged, satisfying the audit requirement for sensitive business sessions.

In practice: A small quality-of-life plugin that solves a real annoyance. The display-only design is thoughtful.

For: Claude (Axion). Claude Code plugin that post-processes every Claude response locally via Ollama; display-only, does not modify the saved transcript.

Security4

Quality3

Auditability3

Useful to you3

Useful to community3

Buildable now4

Hermes1

Verdict: watch. Score 21. Useful for YY's sharing and review workflows but not core infrastructure. Pending marketplace review as of 2026-08-10; install once approved and evaluate over a week before keeping permanently.

Install: pending Anthropic marketplace review; run claude plugins install claudish-to-english once approved. Requires Ollama installed locally with gemma4:26b-mlx or equivalent. Not currently in enabledPlugins. Review before enabling. Copy

Source · Published 2026-08-10 by gvzdv, pending Anthropic marketplace review as of publish date, Ollama local inference confirmed as default, gemma4:26b-mlx default model documented, Anthropic API and OpenAI-compatible endpoints supported as alternatives.

#16

newwatchBoth runtimes22 / 35

Memoria repo

Secure memory management for AI agents: enforces data integrity at write time, reduces hallucinations by restricting recall to verified entries, and maintains consistent long-term context across sessions.

What it does for you: Memoria's core value is write-time integrity enforcement: every memory entry is validated before storage, which prevents the class of errors where a hallucinated fact gets stored and then recalled as truth in future sessions. For MCL agent workflows that generate structured data about clients and campaigns, preventing poisoned memory at the write boundary is more important than retrieval speed. The integrity model is complementary to agentmemory rather than a replacement.

In practice: Security-first in a space that usually optimises for recall speed first. Smaller community but the right instinct for production business data.

For: Both runtimes. Python library from matrixorigin; callable from any runtime that supports Python imports or HTTP endpoints.

Security5

Quality3

Auditability3

Useful to you3

Useful to community2

Buildable now3

Hermes3

Verdict: watch. Score 21. Limited community traction and fewer published benchmarks make quality hard to verify independently. Monitor for peer-reviewed evaluations before adopting for MCL-sensitive data.

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

Source · Repository at matrixorigin/Memoria confirmed live on GitHub, 3 stated design goals in README: data integrity enforcement, hallucination reduction, and consistent long-term context; Python implementation, Apache 2.0 or MIT license per repo metadata.

#17

watchClaude (Axion)25 / 35

awesome-claude-code-toolkit repo

Comprehensive Claude Code resource toolkit: 135 agents, 35 curated skills, 42 commands, 176 plus plugins, 20 hooks, 15 rules, 7 templates, 14 MCP configs, 26 companion apps, and 52 ecosystem entries. Updated with 13 new agent skills on 2026-08-23.

What it does for you: This is the fastest way to survey what is possible with Claude Code without reading 200 separate GitHub repos. The 176 plus plugins and 14 MCP configs sections are directly actionable: scan for anything in the Axion category (scheduling, memory, Git, business tools) and compare with what is already installed. The 2026-08-23 update adds 13 agent skills not catalogued in the hesreallyhim list.

In practice: More comprehensive than the canonical awesome-claude-code list but less curated. Useful as a discovery layer rather than a quality signal.

For: Claude (Axion). Curated collection of Claude Code-native resources: agents, skills, commands, plugins, hooks, and MCP configs all targeting Claude Code.

Security4

Quality3

Auditability3

Useful to you4

Useful to community4

Buildable now5

Hermes2

Verdict: watch. Score 23. A navigation tool rather than a buildable artefact; value is in discovery. Scan the plugins and MCP config sections for Axion-relevant additions after each weekly update.

Build #17 awesome-claude-code-toolkit: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/rohitg00/awesome-claude-code-toolkit. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Repository at rohitg00/awesome-claude-code-toolkit confirmed live, 135 agents and 176 plus plugins count confirmed in README, 13 agent skills added 2026-08-23 per git history, 35 curated skills and 14 MCP configs documented.

#18

newwatchClaude (Axion)21 / 35

awesome-claude-code-plugins repo

Curated list of Claude Code plugins: slash commands, subagents, MCP servers, and hooks. Maintained by the ccplugins community organisation. Distinct from the hesreallyhim general awesome-claude-code list in its focus on the plugin system specifically.

What it does for you: This list is scoped tighter than the broader awesome-claude-code lists: every entry is a plugin-system artefact, which means anything found here can be installed via claude plugins install rather than requiring manual repo setup. For YY's enabledPlugins maintenance, this is the most targeted discovery source for new installable Claude Code capabilities.

In practice: Focused and installation-ready. The plugin-system scope makes it more immediately actionable than the broader awesome lists.

For: Claude (Axion). Curated directory of Claude Code slash commands, subagents, MCP servers, and hooks; all entries target Claude Code natively.

Security4

Quality2

Auditability2

Useful to you4

Useful to community3

Buildable now5

Hermes1

Verdict: watch. Score 20. Discovery resource rather than a buildable artefact. Check monthly for new entries that overlap with Axion's active tool categories.

Build #18 awesome-claude-code-plugins: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/ccplugins/awesome-claude-code-plugins. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Repository at ccplugins/awesome-claude-code-plugins confirmed live on GitHub, 4 categories of Claude Code plugin covered: slash commands, subagents, MCP servers, and hooks per repo description; maintained by ccplugins community organisation.

#19

watchClaude (Axion)21 / 35

quemsah/awesome-claude-plugins repo

Automated collection of Claude Code plugin adoption metrics across GitHub repositories using n8n workflows. Tracks which plugins are being starred, forked, and installed, and surfaces velocity and trend signals.

What it does for you: Rather than reading every plugin repo individually, this tool tells you which Claude Code plugins are gaining traction right now based on real adoption signals. For YY's daily scout process, a plugin that appears here with strong velocity is a higher-confidence scout candidate than one found only in a curated list. The n8n-based data collection is also a template for building similar adoption trackers for other Axion-relevant tool categories.

In practice: Meta-tool that makes plugin discovery data-driven rather than list-dependent. Small but useful as a signal amplifier.

For: Claude (Axion). Analytics tool tracking Claude Code plugin adoption metrics via n8n workflows; outputs are informational and target the Claude Code ecosystem.

Security4

Quality2

Auditability2

Useful to you4

Useful to community3

Buildable now4

Hermes2

Verdict: watch. Score 20. Informational output rather than a directly buildable artefact. Check weekly during the scout run to surface high-velocity plugins before they appear in curated lists.

Build #19 quemsah/awesome-claude-plugins: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/quemsah/awesome-claude-plugins. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Repository at quemsah/awesome-claude-plugins confirmed live on GitHub, n8n-based automation for adoption metric collection stated in repo description, targets GitHub repository metrics for Claude Code plugins.

#20

watchBoth runtimes23 / 35

Mastra repo

TypeScript-first agent framework ranked alongside LangGraph, CrewAI, and n8n in August 2026 framework comparisons. Provides a structured approach to multi-agent orchestration with a TypeScript-native type system.

What it does for you: Mastra occupies the TypeScript niche that LangGraph and CrewAI leave largely unfilled: if any Axion or MCL tooling runs on a Node.js backend, Mastra gives it a typed, framework-native agent layer without bridging to Python. For a stack where most scripts are Python but web tooling (voice-to-content, dashboards) is JavaScript, Mastra could close the gap without requiring a language switch.

In practice: Newer and less proven than LangGraph but a genuine TypeScript-native option rather than a Python port. Watch for community momentum.

For: Both runtimes. TypeScript-first agent framework; callable from Node.js environments in Axion or as a Hermes-accessible HTTP service.

Security4

Quality3

Auditability3

Useful to you3

Useful to community3

Buildable now4

Hermes3

Verdict: watch. Score 22. Appears in top-framework comparisons but lacks specific star count or release date evidence in today's scout run. Monitor for a major release or benchmark before investing time in evaluation.

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

Source · Listed in August 2026 framework comparisons at aimagicx.com and talki-app.fr alongside LangGraph, CrewAI, and n8n, TypeScript-first design confirmed in comparison articles, GitHub repository at mastra-ai/mastra referenced in comparisons.