Daily AI Implementation Scout Council

2026-08-13. Top pick: #1 anthropics/claude-agent-sdk-python. 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

anthropics/claude-agent-sdk-python repo

Official Anthropic Python SDK for building agents that run on Claude Code. Ships structured outputs (agents return validated JSON matching a schema), automatic fallback model handling, a ResultMessage terminal_reason field (shows exactly why the agent loop ended), typed model_usage as dict[str, ModelUsage], SDK beta feature flags including the 1M context window beta, and build scripts hardened against command injection via version validation. Claude Code is bundled in the package by default. Active v0.2.x release track. Apache-2.0 license. 7,700 stars, 1,200 forks.

What it does for you: This is the sanctioned Python path for building Hermes agents that run Claude Code headlessly, replacing ad-hoc CLI subprocess calls with a typed, versioned SDK contract. Structured outputs eliminate the JSON parsing layer that currently sits between Hermes and every tool result it expects to be machine-readable. The terminal_reason field surfaces the exact cause when an agent loop exits early, making debugging silent Hermes failures orders of magnitude faster than reading raw logs.

In practice: 7,700 stars and 1,200 forks for the official Anthropic agent SDK is strong adoption. The feature set matches what a production Hermes pipeline needs: typed outputs, fallback model safety, and observable loop termination. The hardened build scripts are a meaningful supply-chain hygiene signal from the Anthropic team. First appearance in this newsletter despite being the foundation-level SDK for the Axion agent stack.

For: Both runtimes. Official Anthropic Python Agent SDK callable from Axion scripts, Hermes agent loops, and Claude Code sessions.

Security5
Quality5
Auditability5
Useful to you5
Useful to community4
Buildable now4
Hermes5

Verdict: build now. NEW. First appearance in this newsletter. 7,700 stars, 1,200 forks, Apache-2.0. Official Anthropic Python Agent SDK: structured outputs, typed model_usage, terminal_reason, fallback model, command-injection-hardened build scripts. This is the correct foundation for all new Hermes agent builds and should replace any ad-hoc Claude CLI subprocess integration currently in Axion.

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

Source · 7,700 stars, 1,200 forks confirmed from GitHub search results on 2026-08-13. Active v0.2.x release series. Features: structured outputs, fallback model, terminal_reason, model_usage TypedDict, SDK beta flags including 1M context window, build scripts hardened with version validation. Apache-2.0 license. Official Anthropic-maintained repository.

#2
newbuild nowClaude (Axion)31 / 35

anthropics/knowledge-work-plugins plugin

Anthropic's official open-source plugin pack for knowledge workers, compatible with Claude Cowork and Claude Code. Eleven plugins organized by job function: Productivity (task management, workplace memory, visual dashboard), Data (SQL, visualizations, interactive dashboards, analysis validation), Operations (vendor management, process documentation, change management, compliance tracking, resource planning), Design (design critique, system management, UX writing, accessibility, developer handoff), and Legal (compliance). Each plugin bundles skills, connectors, slash commands, and sub-agents for its domain. Apache-2.0 license. 23,000 stars, 2,800 forks. Last updated 2026-08-12.

What it does for you: The Data plugin turns Claude Code into a full SQL analyst and visualization builder, directly applicable to Axion dashboard work, MCL performance data analysis, and here.now report generation. The Operations plugin covers vendor management, capacity planning, and process documentation: the exact tasks that currently generate ad-hoc Axion scripts. Installing both replaces custom skill development time with a vetted official implementation that Anthropic maintains. The Design plugin's developer handoff capability closes the gap between here.now UI builds and team documentation.

In practice: 23,000 stars and 2,800 forks for an official Anthropic plugin release is the strongest community signal of any plugin in this newsletter's history. The multi-plugin structure means YY can install only the domains relevant to MCL work (Data and Operations first, Design second) without enabling everything at once. Apache-2.0 license and Anthropic authorship remove the quality and security uncertainty that typically gates plugin adoption.

For: Claude (Axion). Claude Code and Claude Cowork plugins; not callable from the Hermes Python runtime.

Security5
Quality5
Auditability5
Useful to you5
Useful to community5
Buildable now4
Hermes2

Verdict: build now. NEW. First appearance in this newsletter. 23,000 stars, 2,800 forks, Apache-2.0. Official Anthropic plugin pack with 11 job-function plugins. The Data plugin (SQL, dashboards) and Operations plugin (vendor, process, compliance) are directly applicable to Axion and MCL operations today. Install and evaluate Data and Operations first before enabling the full pack.

Install: enable individual plugins from anthropics/knowledge-work-plugins via Claude Code plugin settings (Apache-2.0, official Anthropic release, updated 2026-08-12). Start with Data and Operations plugins. NOT currently confirmed in YY's enabledPlugins. Review each plugin's data access scope before enabling.

Source · 23,000 stars, 2,800 forks confirmed from GitHub search results on 2026-08-13. 11 plugins confirmed: Productivity, Data, Operations, Design, Legal. Apache-2.0 license. Last repository update 2026-08-12. Official Anthropic-maintained repository. First appearance in this newsletter.

#3
test firstBoth runtimes30 / 35

pydantic/pydantic-ai v2.28.0 repo

Type-safe Python agent framework. New minor release v2.28.0 on 2026-08-11, advancing from v2.27.1. Key additions: realtime speech-to-speech support via Agent.realtime(), a high-severity security patch for the development web chat UI (the endpoint now requires Content-Type: application/json and rejects all other requests before body parsing, closing a request-smuggling-class vulnerability), Cerebras optional dependency group, and fixes for Bedrock cache point attachment, concurrent stream cancellation handling, and realtime reconnection state management.

What it does for you: The high-severity security patch is the mandatory upgrade trigger for any Hermes component running the pydantic-ai development web chat UI. The Agent.realtime() speech-to-speech interface opens a path for voice-driven MCL coaching interactions: a coaching session could use Hermes to drive a realtime Claude conversation without a separate audio pipeline. Upgrade Hermes agents from v2.27.1 to v2.28.0 first, then evaluate Agent.realtime() as a future MCL voice interface layer.

In practice: Minor version advance with a genuine security fix is a quality signal. The realtime speech-to-speech addition is large enough to warrant its own evaluation pass before adopting in a coaching context. The Cerebras addition is provider-specific and not relevant for Anthropic-first stacks. The Bedrock and stream cancellation fixes are stability improvements relevant for high-frequency Hermes runs.

For: Both runtimes. Python agent framework callable from Axion scripts and Hermes agent loops.

Security4
Quality5
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes4

Verdict: test first. New minor release v2.28.0 on 2026-08-11, advancing from v2.27.1. Mandatory security patch for the development web chat UI (high-severity, Content-Type validation added). Agent.realtime() speech-to-speech is a new capability worth evaluating for MCL coaching voice interfaces. Upgrade from v2.27.1, verify the security patch applies to any Hermes dev environment running the chat UI, then test Agent.realtime() on a non-production session.

Build #3 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 · v2.28.0 released 2026-08-11. Previous version v2.27.1 (tracked in yesterday's issue). Security patch: high-severity vulnerability in development web chat UI, Content-Type: application/json validation added. New feature: Agent.realtime() for speech-to-speech. Cerebras optional dependency added. pip install pydantic-ai==2.28.0. Official Pydantic team release.

#4
test firstBoth runtimes30 / 35

BerriAI/litellm v1.96.0 repo

Unified LLM routing layer stable at v1.96.0, released 2026-08-10. Preview track at v1.97.0-rc.1. Docker images are cosign-signed from a pinned commit hash for supply chain verification.

What it does for you: Hermes and Axion scripts routing tool calls through LiteLLM should be upgraded to v1.96.0 if not already done following yesterday's action. No new stable release today. The v1.97.0 stable release is in the preview pipeline, so completing the v1.96.0 upgrade now avoids a two-step cycle.

In practice: No new stable release today. Continued as an open test-first action from yesterday. The v1.97.0-rc.1 in the preview track signals that a v1.97.0 stable is coming soon.

For: Both runtimes. Python LLM router callable from Axion scripts and Hermes agent loops.

Security4
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes5

Verdict: test first. Re-listed with open upgrade action from yesterday. Stable at v1.96.0 (released 2026-08-10), preview at v1.97.0-rc.1. No new stable release today. Test first: upgrade Hermes from any pre-v1.96.0 version and verify tool call compatibility before the v1.97.0 stable creates a second upgrade cycle.

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

Source · v1.96.0 stable released 2026-08-10. Preview track at v1.97.0-rc.1 confirmed. Docker images cosign-signed from pinned commit hash. pip install litellm==1.96.0. Supports 100-plus LLM providers. First listed 2026-08-12.

#5
test firstClaude (Axion)30 / 35

JuliusBrussee/caveman plugin

Claude Code plugin that compresses prompts and code into a compact representation, reducing token usage significantly across long sessions. 61,000-plus stars.

What it does for you: Reduces token burn on every Axion session where large codebases or brain content are loaded into context. The AI Scout newsletter sessions, brain ingest sessions, and MCL coaching content sessions are the highest-value targets for compression. Listed for the third consecutive day and still not in YY's enabledPlugins.

In practice: 61,000-plus stars is the highest community adoption of any plugin tracked in this newsletter. The gap between adoption signal and absent enablement has now persisted across three consecutive issues. The install step is one action.

For: Claude (Axion). Claude Code plugin for token compression; not applicable to the Hermes Python runtime.

Security4
Quality5
Auditability4
Useful to you4
Useful to community5
Buildable now5
Hermes3

Verdict: test first. Re-listed for the third consecutive day. 61,000-plus stars, highest community adoption of any listed plugin. Not yet in YY's enabledPlugins. Test first to verify compression does not mangle skill file syntax, structured JSON outputs, or brain markdown before committing to it for all sessions.

Install: enable JuliusBrussee/caveman in Claude Code plugin settings. NOT currently in YY's enabledPlugins. Review output quality on structured files before enabling session-wide.

Source · 61,000-plus GitHub stars. Claude Code plugin, installable from the marketplace. Not confirmed in YY's enabledPlugins as of 2026-08-13. First listed 2026-08-10.

#6
test firstBoth runtimes29 / 35

affaan-m/agentshield repo

Runtime security scanner for AI agents with 25-plus detection rules: catches prompt injection, data exfiltration, and tool misuse before the agent acts on a tool call.

What it does for you: Adds a defense layer to Hermes and Axion agents at the tool call boundary, catching prompt injection attempts and exfiltration before they reach external systems. Pairs with deepteam at rank 8: deepteam generates red team attacks offline, agentshield blocks them at runtime in production. The pairing remains an open action from 2026-08-11.

In practice: The runtime-guard model at the tool call boundary is the correct architectural approach for agent security. Re-listed alongside deepteam to keep the pairing visible until both are integrated.

For: Both runtimes. Python runtime guard integrating into any Python agent pipeline including Hermes tool call loops.

Security5
Quality4
Auditability4
Useful to you4
Useful to community4
Buildable now4
Hermes4

Verdict: test first. Re-listed to maintain the deepteam and agentshield pairing as an open action from 2026-08-11. Runtime security at the tool call boundary complements deepteam's offline attack catalog. No new release today.

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

Source · 224 forks. 25-plus detection rules for agent security. pip install agentshield. MIT license. First listed 2026-08-10.

#7
test firstBoth runtimes29 / 35

semantica-agi/semantica v0.6.5 repo

Graph-native infrastructure for context and accountable AI systems. v0.6.5 is the current release, closing 6 externally-reported CVEs including 2 Critical vulnerabilities (missing-authentication in the Explorer API and Cypher-injection in triplet store backends). Also added the Oxigraph embedded backend, full PROV-O trust and spec completeness for ProvenanceManager, and the Altair Anzo triplet store backend. No new release confirmed as of 2026-08-13.

What it does for you: The build now action from yesterday was to upgrade to v0.6.5 to close the 2 Critical CVEs. If that upgrade was not completed, it remains the highest-priority action on this item. No new release today means the v0.6.5 actions (upgrade, audit code written against the Explorer API before the upgrade, test the Oxigraph embedded backend in staging) are still the open tasks before the graphify integration can proceed.

In practice: No new release today. Score adjusted downward from yesterday's security-emergency rating: the urgent upgrade trigger has passed (v0.6.5 was released 2026-08-11) and the action should already be complete. Continued as test first to keep the open upgrade action visible.

For: Both runtimes. Python pip package callable from any Python runtime including Hermes.

Security4
Quality5
Auditability5
Useful to you4
Useful to community4
Buildable now4
Hermes3

Verdict: test first. No new release found as of 2026-08-13. Downscored from yesterday's emergency build now (33) because the Critical CVE upgrade trigger has passed; the v0.6.5 upgrade should already be complete. test first remains until the Oxigraph backend and PROV-O staging tests are confirmed done, unblocking the graphify integration decision.

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

Source · v0.6.5 released 2026-08-11. No new release confirmed as of 2026-08-13. 6 CVEs closed in v0.6.5 including 2 Critical. MIT license. First listed 2026-08-11 as rank 2, rank 1 on 2026-08-12.

#8
test firstBoth runtimes28 / 35

confident-ai/deepteam repo

Red team framework for LLMs and AI agents with 50-plus vulnerability types including prompt injection, bias, PII leakage, and jailbreaks. MIT license, pip-installable. Confirmed as one of the best-maintained open-source LLM red teaming frameworks in 2026.

What it does for you: Lets YY run adversarial attack suites against Axion and Hermes agents before each new skill or agent goes live, surfacing prompt injection and data leakage vectors before real users can exploit them. Pairs with agentshield as a complete security posture: deepteam generates the attacks offline, agentshield blocks them at runtime in production. The evaluation pipeline setup and attack type triage for MCL-specific data remain the open action items.

In practice: The 50-plus attack type catalog needs manual triage for MCL-specific risk vectors before this becomes operationally useful rather than noisy. The NIST AgentDojo-Inspect result (81% task-hijack rate on novel attacks vs 11% prior baselines, confirmed today via the security-eval lens) is a reminder that attack surface evaluations need continuous updating.

For: Both runtimes. Standalone Python testing tool; runs against any LLM endpoint including Hermes.

Security3
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes4

Verdict: test first. Re-listed with open action from 2026-08-11: evaluation pipeline setup and attack type triage for MCL-specific data are not yet complete. test first remains correct until the triage step is done. No new release today.

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

Source · 2,416 stars, 390 forks as of last confirmed count. pip install deepteam. MIT license. 50-plus vulnerability types documented. Confirmed as a top-maintained open-source LLM red teaming framework in 2026 via security-eval lens. First listed 2026-08-11.

#9
test firstClaude (Axion)28 / 35

AgriciDaniel/claude-obsidian v1.7 DragonScale repo

Self-organizing AI second brain for Obsidian and Claude Code. v1.6.0 ships DragonScale, a four-mechanism opt-in memory layer for LLM wiki vaults: fold rollups, deterministic page addresses, semantic tiling lint, and boundary-first autoresearch. v1.7 adds the Compound Vault model for multi-vault architectures. Based on the Karpathy LLM Wiki pattern. MIT license. 10,752 stars as of last confirmed count.

What it does for you: DragonScale's deterministic page addresses and boundary-first autoresearch directly address the YY brain's retrieval scaling problem: when brain content grows past 500 pages, these mechanisms provide stable links and autonomous cross-document research that the current grep-first workflow cannot. Compound Vault in v1.7 enables the MCL brain junction pattern (one vault structure referencing another) to be formalized with tooling rather than managed as a fragile Windows junction.

In practice: Significant version feature jump from DragonScale (v1.6.0) and Compound Vault (v1.7) is materially new capability, not incremental maintenance. The four DragonScale mechanisms map directly to identified gaps in the YY brain workflow. This is the most directly applicable pattern tool to YY's current knowledge infrastructure.

For: Claude (Axion). Claude Code integration for the ingest, linking, and vault-intelligence pipeline; not a standalone Python runtime component.

Security3
Quality4
Auditability4
Useful to you5
Useful to community4
Buildable now4
Hermes4

Verdict: test first. Materially updated: DragonScale (v1.6.0) adds fold rollups, deterministic page addresses, semantic tiling lint, and boundary-first autoresearch. v1.7 adds Compound Vault for multi-vault architectures. Both features directly address gaps in the YY 2nd Brain workflow. Test first: evaluate DragonScale's deterministic page addresses against the current brain index before adopting as the primary retrieval model.

Build #9 AgriciDaniel/claude-obsidian: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/AgriciDaniel/claude-obsidian. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · DragonScale four-mechanism memory layer confirmed in v1.6.0. Compound Vault confirmed in v1.7. 10,752 stars from last confirmed count on 2026-08-12. MIT license. First listed 2026-08-12 at test first.

#10
test firstBoth runtimes27 / 35

vitali87/code-graph-rag repo

RAG system for monorepos that builds a knowledge graph from multi-language codebases using Tree-sitter and Memgraph, allowing AI agents to query, understand, and edit across the full codebase. Now ships an MCP server integration for direct Claude Code use. Supports Ruby via ast-grep, data-flow tracing with FLOWS_TO taint edges, and C#, Java, C, and Go, bringing language tracing to 10 languages. Python, MIT license. Active push confirmed 2026-08-11.

What it does for you: The MCP server integration means code-graph-rag can now be connected directly to Claude Code as a context source, providing queryable codebase maps without the graphify refresh cycle. The FLOWS_TO taint edges enable data-flow analysis: tracing where a value produced by one tool call reaches another, which is the class of bug hardest to find in Hermes agent pipelines. The 10-language support covers the full Axion TOOL-AND-APP monorepo language mix.

In practice: The MCP server addition is the most important architectural evolution since first listing: it moves code-graph-rag from a standalone Python tool to a Claude Code-native context source. Momentum from the Aug 12 star spike appears to be stabilizing; star count as of today is uncertain and should be verified directly before citing.

For: Both runtimes. Python RAG tool callable from Axion scripts and Hermes agents, now also available as an MCP server.

Security3
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes3

Verdict: test first. Re-listed. Active push on 2026-08-11, confirmed MCP server integration for direct Claude Code use, FLOWS_TO taint edges for data-flow tracing, and 10-language support covering the full Axion stack. Test first: evaluate the MCP server connection against the Axion TOOL-AND-APP monorepo before replacing any graphify workflow.

Build #10 vitali87/code-graph-rag: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/vitali87/code-graph-rag. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · Active push confirmed 2026-08-11. MCP server integration confirmed. 10-language support confirmed (Python, JavaScript, Ruby, C#, Java, C, Go, and others). FLOWS_TO taint edge data-flow tracing confirmed. MIT license. First listed 2026-08-04.

#11
newwatchBoth runtimes26 / 35

neo4j-labs/create-context-graph repo

Interactive CLI scaffolding tool that generates domain-specific context graph applications in under 5 minutes. Choose an industry domain (22 available) and an agent framework (8 options) and the CLI generates a complete FastAPI backend with an AI agent configured for that domain, powered by neo4j-agent-memory v0.4 for multi-turn conversations. Implements a three-memory-type architecture: short-term (conversation history and document content), long-term (entity knowledge graph on the POLE+O model), and reasoning (decision traces with full provenance). Neo4j Labs project.

What it does for you: Provides a production-ready scaffold for any new Hermes sub-agent that needs persistent context memory with entity extraction and decision provenance: the MCL coaching agent, the FSX forum agent, or the Ryan sales coaching sub-agent each map to one of the 22 industry verticals. The three-memory-type architecture (short-term, long-term entity graph, reasoning traces) is a more complete memory model than Hermes currently implements. The 5-minute scaffold time makes evaluating the pattern low-cost.

In practice: Neo4j Labs origin means the graph infrastructure is production-grade but vendor-tied. The 22-vertical, 8-framework design suggests genuine general-purpose intent. The POLE+O entity model (Person, Object, Location, Event, Organization) is a law-enforcement-derived knowledge graph standard that handles real-world entity resolution better than flat entity lists.

For: Both runtimes. CLI tool that scaffolds FastAPI Python backends with the neo4j-agent-memory layer; the generated apps are Python-based and callable from Hermes.

Security3
Quality4
Auditability4
Useful to you4
Useful to community3
Buildable now4
Hermes4

Verdict: watch. NEW. First appearance in this newsletter. Published CLI scaffolding tool from Neo4j Labs. 22 industry verticals, 8 agent frameworks, three-memory-type architecture including entity knowledge graphs and reasoning traces. Watch: evaluate the scaffold output for the MCL coaching domain before committing to Neo4j as the memory backend.

Build #11 neo4j-labs/create-context-graph: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/neo4j-labs/create-context-graph. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · Published CLI tool confirmed. 22 industry verticals confirmed. 8 agent framework options confirmed. Powered by neo4j-agent-memory v0.4. Three-memory-type architecture confirmed: short-term, long-term entity knowledge graph (POLE+O model), reasoning memory with decision traces. Neo4j Labs official repository. First appearance in this newsletter.

#12
newwatchClaude (Axion)25 / 35

anthropics/claude-code (patch 2026-08-12) repo

Claude Code new patch release confirmed as of 2026-08-12. Three fixes: stream idle timeout no longer fires on custom ANTHROPIC_BASE_URL gateways when server keep-alive pings arrive on the wire; claude.ai connectors are no longer falsely marked as needing authorization when the session token is invalid; tool errors are now displayed for tools no longer available locally. Specific version tag beyond v2.1.228 not confirmed from available sources.

What it does for you: The stream idle timeout fix directly affects any Axion or Hermes session routing through a custom ANTHROPIC_BASE_URL gateway such as a LiteLLM proxy: long-running agent tasks that previously timed out mid-run on gateway keep-alive events will now stay connected. The tool error display fix means sessions now surface errors for missing local tools rather than silently skipping them, which is the class of silent failure most likely to produce wrong agent outputs without any visible signal.

In practice: Three targeted fixes with no feature additions. The gateway idle timeout fix and the tool error display fix are both directly relevant to Axion sessions that use custom endpoint routing. Specific version tag not confirmed beyond the patch description.

For: Claude (Axion). Claude Code release; not a Python runtime component.

Security5
Quality4
Auditability4
Useful to you3
Useful to community4
Buildable now3
Hermes2

Verdict: watch. New patch confirmed 2026-08-12. The gateway idle timeout fix and tool error display fix are directly relevant to Axion sessions. Watch: update Claude Code to the latest version in the current session environment; no action needed beyond keeping the installation current.

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

Source · New patch confirmed as of 2026-08-12. Specific fixes: stream idle timeout on ANTHROPIC_BASE_URL gateways, connectors authorization false positive, tool error display for unavailable local tools. Official Anthropic release. Previous tracked version v2.1.228 (2026-08-11).

#13
watchBoth runtimes24 / 35

modelcontextprotocol/python-sdk v2.0.0 repo

Official Python SDK for the Model Context Protocol, stable at v2.0.0. No new release today. The code-graph-rag MCP server (rank 10) and any future Axion custom MCP servers target this SDK.

What it does for you: Any custom MCP server built for Axion or Hermes should target v2.0.0. The code-graph-rag MCP server integration at rank 10 assumes the v2.x protocol. No new capability today, but the rank 10 item makes the v1.x to v2.0.0 migration reminder freshly relevant.

In practice: No new release. Continued tracking keeps the migration requirement visible for any custom server work still on the v1.x track.

For: Both runtimes. Foundation for any MCP server built in Python; both Claude Code and Hermes consume MCP-compatible servers.

Security5
Quality4
Auditability5
Useful to you2
Useful to community4
Buildable now2
Hermes2

Verdict: watch. No new release today. The v2.0.0 baseline continues as the target for custom MCP server builds. Watch: any custom server on v1.x needs a migration plan before integrating with v2.x-targeted tools such as code-graph-rag's new MCP server.

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

Source · v2.0.0 stable release. Last release date 2026-07-28. pip install mcp. Official MCP Foundation repository. Breaking change from v1.x to v2.0.0 documented in release notes.

#14
watchBoth runtimes24 / 35

agno-agi/agno (updated 2026-08-12) repo

Agent framework updated on 2026-08-12 with AntigravityAgent and AntigravityTools for Google's Antigravity API, GeminiInteractions support for Google managed agents (Deep Research and Antigravity), a data labeling cookbook with 18 self-contained workflows covering text, image, audio, video, document, and composed labeling primitives, and path safety enhancements with centralized security checks.

What it does for you: The data labeling cookbook is the most immediately applicable addition: 18 self-contained workflows for labeling across media types covers the MCL content-bank labeling and coaching content annotation use cases without building from scratch. The path safety enhancements harden any Axion script using Agno against path traversal attacks. The Antigravity and Gemini additions are Google-stack specific and not immediately relevant for an Anthropic-first deployment.

In practice: The data labeling cookbook is a meaningful practical addition that extends Agno beyond agent orchestration into structured content processing, which overlaps the MCL content pipeline directly. The Google API additions confirm Agno is expanding provider coverage.

For: Both runtimes. Python agent framework with native semantica integration and updated Google API support.

Security3
Quality4
Auditability3
Useful to you4
Useful to community4
Buildable now3
Hermes3

Verdict: watch. Updated 2026-08-12 with a data labeling cookbook (18 workflows) and path safety enhancements. Watch: evaluate the data labeling cookbook for MCL content-bank annotation use cases before building a custom labeling pipeline in Axion.

Build #14 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 · Updated 2026-08-12 confirmed from search results. AntigravityAgent and AntigravityTools confirmed. GeminiInteractions confirmed. Data labeling cookbook with 18 workflows confirmed. Path safety enhancements confirmed. pip install agno. MIT license.

#15
watchClaude (Axion)23 / 35

anthropics/claude-plugins-official repo

Anthropic-managed official directory for Claude Code plugins. The knowledge-work-plugins release (rank 2 today) is the primary new addition tracked via this lens this issue. A community mirror (anthropics/claude-plugins-community) and the rdmgator12/awesome-claude-plugins directory (132 plugin bundles, community-maintained) are companion discovery sources. Apache-2.0 license.

What it does for you: Tracking this directory is how the newsletter surfaced the knowledge-work-plugins release at rank 2. Continued scanning identifies new plugins before they appear in the Claude Code marketplace UI.

In practice: Active directory. The knowledge-work-plugins addition at rank 2 is the headline find today. Companion resources extend coverage beyond official additions.

For: Claude (Axion). The official Claude Code plugin directory; changes here add or update Claude Code plugins.

Security5
Quality3
Auditability4
Useful to you3
Useful to community4
Buildable now2
Hermes2

Verdict: watch. knowledge-work-plugins (rank 2 today) is the primary find from this lens. Watch: continue scanning for new official additions that surface plugins before the marketplace UI does.

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

Source · anthropics/claude-plugins-official confirmed active. knowledge-work-plugins confirmed: 23,000 stars, 2,800 forks, Apache-2.0, 11 plugins, updated 2026-08-12. anthropics/claude-plugins-community mirror confirmed. rdmgator12/awesome-claude-plugins: 132 plugin bundles confirmed.

#16
watchBoth runtimes22 / 35

karpathy/autoresearch repo

Andrej Karpathy's automated research agent: takes a research question, searches, reads, and synthesizes findings autonomously. 13,300-plus forks.

What it does for you: Provides a reference implementation of a self-directed research loop that YY can adapt for AI scout scouting automation or for the night-shift queue's research sub-tasks. High credibility source due to the author.

In practice: No new releases or commits confirmed today. Continued as a reference implementation for research agent loops.

For: Both runtimes. Python research agent runnable standalone or as a Hermes sub-routine.

Security3
Quality4
Auditability3
Useful to you4
Useful to community4
Buildable now2
Hermes2

Verdict: watch. No new releases or commits today. Continues as a reference implementation for research agent loops. Watch until a stable release or a specific Axion integration point is identified.

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

Source · 13,300-plus forks confirmed. MIT license. By Andrej Karpathy. First listed 2026-08-02.

#17
watchBoth runtimes21 / 35

openai/openai-agents-python repo

Official OpenAI agent SDK. Recent addition: Programmatic Tool Calling via agents.tool.ProgrammaticToolCallingTool, which lets supported OpenAI Responses models generate JavaScript to coordinate eligible tools. Also fixes for agent tool name collisions, persisting max-turns handler output to sessions, and improved streamed runs. Apache-2.0 license.

What it does for you: Programmatic Tool Calling is a novel agent pattern: the model generates JavaScript to coordinate tools rather than a fixed tool-call schema. This is worth studying as a pattern reference for Hermes tool orchestration design, even if the OpenAI-first implementation is not directly adopted.

In practice: No specific new release date confirmed for today. Re-listed for the Programmatic Tool Calling pattern, which is a conceptually novel agent coordination mode. Anthropic-first stack means this is pattern reference, not a build target.

For: Both runtimes. Python agent SDK callable from Axion scripts and Hermes, though optimized for OpenAI models.

Security3
Quality4
Auditability3
Useful to you2
Useful to community4
Buildable now3
Hermes2

Verdict: watch. Re-listed for the Programmatic Tool Calling pattern (JavaScript-driven tool coordination) as a conceptually novel agent orchestration reference. Watch: review the pattern for Hermes architecture comparison. Axion is Anthropic-first and this is not a direct build target.

Build #17 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 · Programmatic Tool Calling feature confirmed via PR 2480. Tool name collision fixes confirmed. Apache-2.0 license. Official OpenAI release. pip install openai-agents.

#18
watchBoth runtimes21 / 35

crewAIInc/crewAI repo

Multi-agent orchestration framework stable at v1.15.14. No new release since 2026-08-08.

What it does for you: Provides a role-based multi-agent pattern that can be adapted for Hermes multi-step workflows where specialist agents hand off work. Useful as a pattern reference even if Axion uses its own orchestration layer.

In practice: Stable and mature. No compelling new feature today. Re-listed to maintain awareness of the multi-agent framework landscape.

For: Both runtimes. Python multi-agent framework callable from Axion scripts and Hermes.

Security3
Quality4
Auditability3
Useful to you3
Useful to community4
Buildable now2
Hermes2

Verdict: watch. No new release since 2026-08-08. Continues as a reference for multi-agent orchestration patterns. Watch until a new major release or a specific Hermes integration case is identified.

Build #18 crewAIInc/crewAI: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/crewAIInc/crewAI. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · v1.15.14 confirmed stable. pip install crewai. Apache-2.0 license. Last release 2026-08-08.

#19
watchBoth runtimes20 / 35

earendil-works/pi repo

TypeScript agent harness with vendor-neutral telemetry contracts (pi-telemetry), unified multi-provider LLM API, agent runtime with tool calling, coding agent CLI, and terminal UI. 5,582 commits across multiple packages.

What it does for you: The pi-telemetry vendor-neutral telemetry contracts (typed schemas with conformance tests for agent observability) are the standalone value: usable from Python via the schema definitions independently of adopting the full TypeScript stack. Useful as a telemetry baseline for Hermes observability design.

In practice: No new release today. TypeScript is the main friction for Axion integration, but the telemetry contracts are language-agnostic schemas. Continued at rank 19 as the telemetry reference item.

For: Both runtimes. TypeScript agent harness; the pi-telemetry contracts are language-agnostic schemas usable from Python.

Security3
Quality3
Auditability3
Useful to you3
Useful to community4
Buildable now2
Hermes2

Verdict: watch. No new release today. Re-listed for the pi-telemetry vendor-neutral telemetry contracts as standalone value, independent of whether the full agent runtime is adopted.

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

Source · 5,582 commits confirmed. Multi-package monorepo with pi-telemetry, pi-ai, pi-agent-core, pi-coding-agent, pi-tui packages. TypeScript. First listed 2026-08-11.

#20
skipClaude (Axion)19 / 35

hesreallyhim/awesome-claude-code repo

Community-curated list of Claude Code resources, plugins, tools, and tips. Several new resource submissions were opened around 2026-08-08 (issues 2469, 2468, 2467, 2466 from contributors including Q00, tonite31, Scalix-World-team, and Opaland, several marked as passing initial validation). The knowledge-work-plugins release at rank 2 will likely drive further entries in the coming days.

What it does for you: Acts as a broad discovery surface for Claude Code plugins and tools not yet on the newsletter radar. The plugin scout lens uses it as one of four checked sources each issue. New submissions from the Aug 8 pipeline may yield actionable items in the next 1 to 2 issues.

In practice: Active submission pipeline as of Aug 8. The knowledge-work-plugins release will likely drive new entries soon. Listed at rank 20 as the discovery backbone of the plugin scout lens.

For: Claude (Axion). Curated discovery list for Claude Code resources; no runtime component.

Security4
Quality3
Auditability4
Useful to you2
Useful to community4
Buildable now1
Hermes1

Verdict: skip. No confirmed new user-contributed entries since 2026-08-08; several submissions in validation pipeline as of that date. Pure reference resource; not a buildable artifact. Included at rank 20 as the plugin scout source checked each issue alongside the three other plugin-scout sources.

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

Source · New resource submissions confirmed around 2026-08-08: issues 2469, 2468, 2467, 2466. Several marked as passing initial validation. NOASSERTION license. Plugin scout checked 2026-08-13; no new entries confirmed beyond the Aug 8 pipeline.