Daily AI Implementation Scout Council

2026-09-03. Top pick: #1 Anthropic SDK Python v1.3.0. Each item is graded on 7 axes; copy a build command to act on it.

Today's ranked top 20

#1
newbuild nowBoth runtimes33 / 35

Anthropic SDK Python v1.3.0 repo

Anthropic's official Python SDK, version 1.3.0, released 2026-09-01. Adds organization compliance settings, user-profile order_by query support, and memory-store plus toolset schema updates to the beta API. Bug fixes include AWS base_url resolution under skip_auth, batch results exposed in GA raw/streaming response wrappers, and multipart filename derivation for file tuples passed without one.

What it does for you: The batch results fix is the most immediately impactful change: any Axion script that polls for async batch output using the GA response wrapper now reliably gets the results field, removing a silent gap that existed since the batch API hit GA. The memory-store schema updates matter for any Axion agent that stores context between sessions through the SDK memory API. Update the version pin in Axion and Hermes environments, run existing agent scripts, and confirm response shapes before treating the upgrade as done.

In practice: A tidy cleanup release that makes the beta memory and toolset APIs more reliable without introducing breaking changes to stable namespaces. The org compliance settings are enterprise-facing, but the batch and multipart fixes land for every caller.

For: Both runtimes. Core Python library for calling Claude; every Axion and Hermes script that calls the API runs through this.

Security5
Quality4
Auditability5
Useful to you5
Useful to community5
Buildable now5
Hermes4

Verdict: build now. Score 33. The batch results fix and memory-store schema update are immediately usable improvements that require no new code: update the pin, run tests, done. The compliance and user-profile features are beta but non-breaking for existing code, and the AWS base_url fix closes a real edge case for any Axion deployment that targets a regional Claude endpoint.

Build #1 Anthropic SDK Python v1.3.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/anthropics/anthropic-sdk-python/releases/tag/v1.3.0. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · v1.3.0 released 2026-09-01 by stainless-app on anthropics/anthropic-sdk-python; full changelog available at github.com/anthropics/anthropic-sdk-python/compare/v1.2.0...v1.3.0; commit b6d1732 fixes AWS base_url resolution; commit cbf9715 adds batch results to GA raw/streaming response wrappers; commit 74080c3 adds external_user_onboarded_at to beta user profiles and replaces relationship field with access_type; commit 429e719 adds org compliance settings and memory-store and toolset schema updates; commit a9f3fb4 fixes multipart filename derivation.

#2
newtest firstClaude (Axion)30 / 35

Claude Code v2.1.258 repo

Claude Code CLI version 2.1.258, current Latest release as of 2026-09-03. Two targeted bug fixes: a macOS 12 Monterey launch failure caused by a regression introduced in v2.1.255, and a remote/scheduled session crash where re-sending a permission approval left an empty user message that the model refused with 'user messages must have non-empty content'.

What it does for you: The scheduled session fix is directly relevant to your setup. The Axion bridge sends permission approvals for unattended scheduled tasks, and this exact crash path was open since v2.1.255. If any bridge-dispatched session failed silently in the last few days, this patch closes the hole. The macOS fix is irrelevant on Windows but confirms the release is a real patch cycle, not a speculative bump.

In practice: Two targeted fixes, not polish. The scheduled session crash is a hard blocker for any unattended agent workflow and its resolution is the most operationally meaningful fix since the v2.1.251 hook changes dropped.

For: Claude (Axion). Claude Code CLI itself; the version running in this session.

Security5
Quality4
Auditability4
Useful to you5
Useful to community5
Buildable now4
Hermes3

Verdict: test first. Score 30. Verify the next unattended bridge session completes without the non-empty content error before relying on automated approval flows again. The patch is a direct fix and no new code is needed.

Build #2 Claude Code v2.1.258: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.258. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · v2.1.258 tagged as Latest on anthropics/claude-code as of 2026-09-03; 2 bugs fixed per release notes: macOS 12 launch regression introduced in v2.1.255, and remote/scheduled session 'user messages must have non-empty content' after permission re-send; 19 community emoji reactions on the release page; yesterday's newsletter covered v2.1.257 (5 patches, regression window open).

#3
test firstBoth runtimes29 / 35

AutoGen python-v0.7.5 repo

Microsoft AutoGen Python framework version 0.7.5, released 2026-09-02. Adds native Anthropic extended thinking mode support, RedisMemory linear memory access, and security warnings that now default code execution to DockerCommandLineCodeExecutor. Also fixes Bedrock streaming with tool usage returning empty arguments, GraphFlow cycle detection state cleanup, Azure AI streaming finish_reason logic, and Redis cache invalidation.

What it does for you: The Anthropic thinking mode addition means any AutoGen agent calling Claude can now request extended reasoning natively, without a custom wrapper. The RedisMemory linear memory option is the more strategically important change: linear memory means AutoGen agents maintain ordered conversation context across sessions rather than a flat key-value store, which maps directly onto how Axion brain ingest agents need to accumulate and retrieve knowledge in sequence. The Docker executor default is a one-time migration if any existing AutoGen script relied on local code execution.

In practice: The Anthropic thinking mode support is the headline feature for this setup. Combined with the Redis linear memory improvement, this release makes AutoGen a more credible alternative to rolling fully custom agent orchestration in Axion.

For: Both runtimes. Multi-provider agent framework; works with Claude via the Anthropic client and with Hermes pipeline scripts via its provider-agnostic core.

Security4
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes4

Verdict: test first. Score 29. Build one AutoGen agent that calls Claude with thinking mode enabled and verify the response shape matches expectations. Redis linear memory requires Redis 7+; test against a local instance before wiring into production Axion agents. The Docker executor default is a breaking change for any existing AutoGen script that uses local code execution.

Build #3 AutoGen python-v0.7.5: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/microsoft/autogen/releases/tag/python-v0.7.5. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · python-v0.7.5 tagged on microsoft/autogen, released 2026-09-02; commit 83afbf5 signed by GitHub (GPG key B5690EEEBB952194); 8 commits to main since release tag; PR #7002 adds thinking mode for Anthropic client (SrikarMannepalli); PR #6972 adds linear memory in RedisMemory (justin-cechmanek); PR #7035 adds security warnings and Docker executor default (ekzhu); PR #6979 fixes Bedrock streaming tool usage returning empty arguments (pawel-dabro); PR #7026 fixes GraphFlow cycle detection (Copilot).

#4
test firstClaude (Axion)29 / 35

GitHub MCP Server v1.11.0 plugin

GitHub's official MCP server, version 1.11.0, still Latest as of 2026-09-03. Adds 7 updated tool endpoints and merges 5 pull requests improving issue search, pull request filtering, and repository context retrieval. Carried forward from Sep 02 with the test-first action still open.

What it does for you: The improved pull request and issue search endpoints are directly useful for the Axion docs git workflow. Not present in YY's enabledPlugins as of 2026-09-03. The test-first window from Sep 02 is still open and narrowing: if no version has been tested against the current Claude Code session by the next issue, this item will be dropped.

In practice: No change since Sep 02. Still the most capable GitHub integration available for Claude Code. The test window is closing.

For: Claude (Axion). MCP server installed into Claude Code; not a standalone service or Hermes component.

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

Verdict: test first. Score 29. Carry forward from Sep 02. Install the update, verify existing GitHub MCP workflows behave correctly, and check for tool endpoint signature changes in PRs #3128 and #3134 before relying on automated workflows.

Install: update the github-mcp-server entry in your Claude Code MCP config to v1.11.0 and restart Claude Code. Review before enabling.

Source · v1.11.0 confirmed as Latest on github/github-mcp-server as of 2026-09-03; 5 PRs merged: #3128, #3134, #3026, #2900, #3147; 7 existing tool endpoints updated; not present in YY's enabledPlugins as of 2026-09-03.

#5
test firstBoth runtimes29 / 35

Pydantic-AI v2.37.0 repo

Pydantic-AI v2.37.0, released 2026-08-31, still Latest as of 2026-09-03. Provides typed agent definitions, structured output validation, and MCP tool integration. This version adds glm-5.3-flash support, fixes AG-UI TEXT_MESSAGE_START emission for tool-calling responses, corrects Vertex vs Gemini API routing by client transport, and fixes durable operation context management.

What it does for you: The AG-UI fix and durable operations fix are the most relevant changes for any Axion agent that uses streaming or long-running operations. The MCP tool integration makes it straightforward to wire existing Axion MCP servers into a typed agent. Carried forward from Sep 02 with the test-first proof-of-concept action still open: no proof of concept has been built yet.

In practice: Still the cleanest typed agent framework for Claude-backed agents. No new version dropped today, which means the Sep 02 action remains pending rather than urgent.

For: Both runtimes. Python agent framework compatible with Claude and Hermes pipelines via its provider-agnostic model abstraction.

Security4
Quality4
Auditability4
Useful to you4
Useful to community5
Buildable now4
Hermes4

Verdict: test first. Score 29. Carry forward from Sep 02. The proof-of-concept action is still open: build one typed agent against one existing Axion task and compare code volume against the current custom approach.

Build #5 Pydantic-AI v2.37.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/pydantic/pydantic-ai/releases/tag/v2.37.0. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · v2.37.0 released 2026-08-31, Latest tag confirmed 2026-09-03; PR #7887 adds glm-5.3-flash and reworks Z.AI test suite onto cassettes (dsfaccini); PR #7699 fixes AG-UI TEXT_MESSAGE_START for tool-calling responses; PR #7280 fixes Vertex vs Gemini routing by client transport; PR #7954 fixes durable operations context management; 4 new contributors in this release.

#6
test firstClaude (Axion)28 / 35

mcp-server-dev plugin

Official Anthropic Claude Code plugin for MCP server development. Provides inline documentation, code completion context, and schema validation for building MCP servers directly inside Claude Code sessions.

What it does for you: You build MCP servers for Axion: the Telegram bridge server, brain ingest endpoints, and planned scout indexer. This plugin injects MCP protocol schema context into coding sessions, catching schema mismatches and surfacing tool definition patterns without a tab-switch to docs. Not present in YY's enabledPlugins as of 2026-09-03; enabling it during the next MCP server build would improve it immediately.

In practice: More targeted than the generic agent-sdk-dev plugin and directly matched to the kind of infrastructure work done daily in Axion. The action is straightforward and reversible.

For: Claude (Axion). Official Anthropic Claude Code plugin; Claude Code environment only.

Security5
Quality4
Auditability4
Useful to you5
Useful to community4
Buildable now3
Hermes3

Verdict: test first. Score 28. Enable in a test session during the next MCP server build and confirm the schema hints are accurate for the protocol version you are targeting before relying on generated boilerplate.

Install: add mcp-server-dev to enabledPlugins in ~/.claude/settings.json and reload Claude Code. Review before enabling.

Source · mcp-server-dev folder confirmed in anthropics/claude-plugins-official/plugins tree, Apache-2.0 license on host repo; plugin directory contains more than 20 official plugin folders as of 2026-09-03 fetch; not present in YY's enabledPlugins as of 2026-09-03.

#7
test firstClaude (Axion)28 / 35

claude-security plugin

Anthropic's official Claude Code security plugin. Surfaces security-relevant context during code review sessions, flagging credential exposure patterns, insecure defaults, and common injection vectors inline.

What it does for you: For Axion and Hermes work you currently run security audits manually or via the gsd-secure-phase skill. This plugin automates the first pass: credential scan, injection pattern detection, and insecure config flags, right inside Claude Code. Not present in YY's enabledPlugins as of 2026-09-03. Carried forward from Sep 02 with the enable action still open.

In practice: Fills an obvious gap in the plugin lineup. Anthropic-signed, installs cleanly, and the action is a one-line settings change.

For: Claude (Axion). Anthropic official Claude Code plugin; Claude Code environment only.

Security5
Quality4
Auditability4
Useful to you5
Useful to community4
Buildable now3
Hermes3

Verdict: test first. Score 28. Enable in a test session and confirm it does not surface false positives on your existing AXION codebase before enabling globally.

Install: add claude-security to enabledPlugins in ~/.claude/settings.json and reload Claude Code. Review before enabling.

Source · claude-security confirmed in anthropics/claude-plugins-official/plugins tree; Apache-2.0 license on host repo; not present in YY's enabledPlugins as of 2026-09-03.

#8
test firstClaude (Axion)28 / 35

agent-sdk-dev plugin

Anthropic's official plugin for Claude Agent SDK development. Provides inline documentation, schema validation, and code completion context for building agents with the Claude Agent SDK.

What it does for you: You build agents directly against the Claude Agent SDK for Axion and Hermes. This plugin injects SDK schema context into coding sessions. Now that SDK v1.3.0 is the recommended update (rank 1 today), the agent-sdk-dev plugin schema hints should reflect the updated memory-store and toolset shapes. Not present in YY's enabledPlugins as of 2026-09-03.

In practice: Purpose-built for the exact workflow you run daily. More relevant today given the SDK v1.3.0 memory-store schema updates landing alongside it.

For: Claude (Axion). Claude Code plugin for agent development assistance; Claude Code environment only.

Security5
Quality4
Auditability4
Useful to you5
Useful to community4
Buildable now3
Hermes3

Verdict: test first. Score 28. Enable alongside the SDK v1.3.0 update to confirm schema hints match the new memory-store and toolset shapes. Test one context-injection plugin at a time to avoid conflicts.

Install: add agent-sdk-dev to enabledPlugins in ~/.claude/settings.json and reload Claude Code. Review before enabling.

Source · agent-sdk-dev confirmed in anthropics/claude-plugins-official/plugins tree; Apache-2.0 license on host repo; not present in YY's enabledPlugins as of 2026-09-03.

#9
test firstBoth runtimes28 / 35

Phoenix v20.5.0 repo

Arize Phoenix v20.5.0, released 2026-09-01. Adds skill tools to the MCP server, a session-level PII detection evaluator, and a trace-replay sidecar for recorded corpora. Still Latest as of 2026-09-03. Carried forward from Sep 02 with the test-first action still open.

What it does for you: The MCP skill tools addition means Phoenix can call Claude Code skills from within its eval loop, enabling automated skill-quality evaluation without custom harness code. The session-level PII detection evaluator is directly applicable to the Telegram bridge, which processes messages from real MCL users. The test-first action from Sep 02 is still open: no proof of concept has been built.

In practice: Three genuinely new capabilities in one release. Remains the most capable open-source eval and observability option for LLM agents that is not already embedded in a vendor's stack.

For: Both runtimes. Open-source eval and observability platform; provider-agnostic, works with Claude and Hermes pipelines.

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

Verdict: test first. Score 28. Test the MCP skill tool integration against a single Axion skill before committing Phoenix as an eval layer. The scout-newsletter skill is a good first target since its output is already structured and verifiable.

Build #9 Phoenix v20.5.0: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/Arize-ai/phoenix/releases/tag/arize-phoenix-v20.5.0. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · arize-phoenix-v20.5.0 released 2026-09-01 on Arize-ai/phoenix; Latest tag confirmed 2026-09-03; 3 new features: MCP skill tools, session-level PII detection evaluator, trace-replay sidecar; MIT license; Phoenix GitHub repo has 3,600+ stars.

#10
watchClaude (Axion)26 / 35

skill-creator plugin

Official Anthropic plugin for creating and maintaining Claude Code skills. Provides inline guidance, scaffolding templates, and validation for the skill authoring workflow within Claude Code.

What it does for you: Already enabled in YY's settings as skill-creator@claude-plugins-official. This plugin is actively available in every Claude Code session. It surfaces Anthropic's official skill authoring conventions inline, which is relevant whenever you author or update a skill in AXION\Skills. No install action needed; the value comes from actively using it during the next skill build.

In practice: Already live. The outstanding action is to use it intentionally during the next skill authoring session and compare its scaffolding conventions against the AXION\Skills format you have standardized on.

For: Claude (Axion). Official Anthropic Claude Code plugin for skill authoring; Claude Code environment only.

Security4
Quality4
Auditability4
Useful to you4
Useful to community4
Buildable now3
Hermes3

Verdict: watch. Score 26. Already enabled; no install action needed. Monitor whether the scaffold templates match AXION\Skills conventions and raise a flag if they conflict, since a mismatch could introduce inconsistent skill structures.

Install: already enabled as skill-creator@claude-plugins-official in YY's settings.json. No install action needed. Review before enabling.

Source · skill-creator@claude-plugins-official: true confirmed in YY's ~/.claude/settings.json enabledPlugins as of 2026-09-03; Apache-2.0 license on host repo anthropics/claude-plugins-official; plugin directory confirmed containing 20+ plugin folders.

#11
watchBoth runtimes25 / 35

Composio awesome-claude-skills repo

Composio's curated collection of 30 Claude Code skills plus 832 SaaS automation templates wired via the Composio SDK. Includes skills for artifacts-builder, brand-guidelines, canvas-design, changelog-generator, and competitive-ads-extractor. Each skill connects Claude to 1,000+ SaaS apps through a single connect-apps:setup command.

What it does for you: The competitive-ads-extractor skill is immediately relevant for MCL: it pulls ad creative and copy from competitor campaigns, which maps onto the content research YY already does manually before scripting angles. The changelog-generator skill could automate the AXION skill changelog workflow. The Composio SDK handles auth and rate limiting for the SaaS connections, removing the boilerplate that currently lives in custom Axion bridge code.

In practice: First appearance in this newsletter. The star count (49,300+) signals broad community adoption. The risk is Composio's proprietary SDK layer adding a vendor dependency on top of the Claude Code skill pattern.

For: Both runtimes. Language-agnostic skill collection and SaaS automation templates; usable in both Claude Code sessions and Hermes pipelines via the Composio SDK.

Security3
Quality3
Auditability3
Useful to you4
Useful to community5
Buildable now4
Hermes3

Verdict: watch. Score 25. Review the competitive-ads-extractor and changelog-generator skills against your current manual workflows before committing to the Composio SDK dependency. The templates are the value; the SDK is the cost.

Build #11 Composio awesome-claude-skills: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/ComposioHQ/awesome-claude-skills. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · github.com/ComposioHQ/awesome-claude-skills confirmed live; 49,300+ stars on repo; 30 curated skills listed including artifacts-builder, brand-guidelines, canvas-design, changelog-generator, competitive-ads-extractor; 832 SaaS automation templates; connect-apps:setup installs via CLI; free API key at dashboard.composio.dev.

#12
watchClaude (Axion)25 / 35

VoltAgent awesome-claude-code-subagents repo

VoltAgent's collection of 100+ specialized Claude Code subagents covering backend, frontend, mobile, and language-specific development. The voltagent-core-dev and voltagent-lang plugin sets are already enabled in YY's settings. Two README commits landed on Sep 01 and Sep 02, 2026, suggesting new agents were added this week.

What it does for you: voltagent-core-dev and voltagent-lang are already active in every Claude Code session, providing backend-developer, frontend-developer, fullstack-developer, api-designer, and 20+ language specialists as callable subagents. The Sep 1-2 README commits may include new specialist subagents. Worth checking the current catalog for additions relevant to current Axion build work.

In practice: Already live and providing value. The Sept 1-2 activity keeps this relevant for checking whether new specialist agents landed that match upcoming Axion build needs.

For: Claude (Axion). Claude Code subagent definitions installed into Claude Code; voltagent-core-dev and voltagent-lang are already enabled in YY's settings.

Security3
Quality3
Auditability3
Useful to you4
Useful to community5
Buildable now4
Hermes3

Verdict: watch. Score 25. Already enabled via voltagent-core-dev@voltagent-subagents and voltagent-lang@voltagent-subagents in settings.json. Monitor the README for new agent additions that match upcoming Axion build priorities.

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

Source · VoltAgent/awesome-claude-code-subagents confirmed 15,600+ stars; 2 README commits on 2026-09-01 (d1fd754) and 2026-09-02 (009544a, 045bdea); voltagent-core-dev@voltagent-subagents: true and voltagent-lang@voltagent-subagents: true confirmed in YY's settings.json; subagent-catalog tool provides /subagent-catalog:search and /subagent-catalog:list commands.

#13
newwatchBoth runtimes24 / 35

LiteLLM v1.99.1 repo

LiteLLM v1.99.1, released as a Docker-only build. No PyPI package exists for this version; pip install litellm==1.99.1 will not resolve and the PyPI release stays at v1.99.0. Carries one fix on top of v1.99.0: OpenTelemetry v2 spans now emit cache token counts (gen_ai.usage.cache_creation.input_tokens and gen_ai.usage.cache_read.input_tokens) alongside the cache cost that was already reported.

What it does for you: If any Axion or Hermes pipeline computes spend from OTel token counts rather than from LiteLLM's own cost fields, prompt-caching workloads were previously under-counted by the full amount of cache creation and cache read tokens. The scout newsletter itself could benefit from accurate cache token tracking if it is ever instrumented with OTel. This is a Docker-only fix, so the action is only relevant if you run LiteLLM as a container rather than as a pip package.

In practice: A narrow but technically correct fix. The Docker-only delivery is unusual and limits the audience to container deployments. Worth tracking for anyone running LiteLLM as a proxy.

For: Both runtimes. Provider-agnostic LLM proxy and SDK; works with Claude and any Hermes pipeline that routes through LiteLLM.

Security4
Quality3
Auditability4
Useful to you3
Useful to community4
Buildable now3
Hermes3

Verdict: watch. Score 24. Docker-only release with no PyPI package. Monitor for a PyPI equivalent if you run LiteLLM as a pip dependency. If you run LiteLLM as a container, pull the v1.99.1 image to get the OTel cache token fix.

Build #13 LiteLLM v1.99.1: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/BerriAI/litellm/releases/tag/v1.99.1. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · v1.99.1 released on BerriAI/litellm as Docker-only; confirmed Latest image tag as of 2026-09-03; images: ghcr.io/berriai/litellm:1.99.1 and docker.io/litellm/litellm:1.99.1; 1 fix over v1.99.0: OTel v2 spans emit gen_ai.usage.cache_creation.input_tokens and gen_ai.usage.cache_read.input_tokens; PyPI stays at v1.99.0; release notes explicitly state pip install litellm==1.99.1 will not resolve.

#14
watchStandalone tool23 / 35

Khoj repo

Khoj is a self-hostable AI second brain with 37,000+ stars on GitHub. It provides web and document search, custom agent creation, scheduled automations, and deep research across personal knowledge bases. It supports multiple LLM backends including Claude, GPT, Gemini, Llama, Qwen, and Mistral. Trending today at 105 stars gained in a single day.

What it does for you: Khoj overlaps conceptually with the YY-2nd Brain vault and the Axion brain ingest pipeline. Running Khoj alongside the existing brain would create redundancy, but its scheduled automations and web search agents could offload research tasks that currently require manual Axion sessions. The multi-LLM backend means it could run on a local model for cost-sensitive tasks while still calling Claude for high-quality summaries.

In practice: The concept is very close to what Axion already does, but Khoj ships as a complete product with a polished interface. The risk is overlap and maintenance complexity rather than a gap in capability.

For: Standalone tool. Self-hostable AI second brain that works with multiple LLM backends; not a Claude Code or Hermes native integration but can call Claude as one of its providers.

Security3
Quality3
Auditability3
Useful to you4
Useful to community5
Buildable now3
Hermes2

Verdict: watch. Score 23. First appearance in the newsletter. The overlap with existing Axion brain infrastructure makes it a watch rather than an action item. Monitor for specific features that do not exist in the current stack, particularly the scheduled research automations.

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

Source · github.com/khoj-ai/khoj confirmed 37,000 stars and 2,448 forks as of 2026-09-03; trending at 105 stars gained today; supports Claude, GPT, Gemini, Llama, Qwen, Mistral backends; built-in scheduled automations and deep research; MIT license.

#15
watchClaude (Axion)23 / 35

claude-md-management plugin

Official Anthropic plugin for managing and synchronizing CLAUDE.md files across Claude Code projects. Provides inline guidance and validation for CLAUDE.md structure, instruction quality, and cross-project consistency.

What it does for you: You maintain multiple CLAUDE.md files: one global at ~/.claude/CLAUDE.md, one at AXION, and separate contracts in YY-2nd Brain and MCL 2nd Brain. This plugin could surface inconsistencies across those files and validate that instructions follow Anthropic's recommended structure. Not present in YY's enabledPlugins as of 2026-09-03.

In practice: A natural companion to the skill-creator plugin already enabled. The multi-CLAUDE.md setup in Axion is exactly the kind of complexity this plugin was designed for.

For: Claude (Axion). Official Anthropic Claude Code plugin; Claude Code environment only.

Security4
Quality3
Auditability4
Useful to you3
Useful to community4
Buildable now3
Hermes2

Verdict: watch. Score 23. Not yet tested against the Axion CLAUDE.md setup. Enable in a test session and verify that validation rules do not conflict with existing CLAUDE.md structures, particularly the brain contract files.

Install: add claude-md-management to enabledPlugins in ~/.claude/settings.json and reload Claude Code. Test against one project CLAUDE.md before enabling globally. Review before enabling.

Source · claude-md-management folder confirmed in anthropics/claude-plugins-official/plugins tree as of 2026-09-03 fetch; Apache-2.0 license on host repo; not present in YY's enabledPlugins as of 2026-09-03.

#16
newwatchBoth runtimes22 / 35

AgentJudgeBench repo

AgentJudgeBench is a multi-difficulty benchmark from ServiceNow-AI for evaluating LLM-as-judge performance on agentic tool-calling tasks. It tests how well a judge model scores agent outputs across easy, medium, and hard tool-calling scenarios, exposing systematic biases in judge model quality that are invisible in single-difficulty benchmarks.

What it does for you: Any Axion or Hermes eval pipeline that uses Claude as a judge to score agent outputs (such as the scout newsletter quality eval or the brain ingest summarizer) is subject to the systematic biases this benchmark measures. Running AgentJudgeBench against your current judge setup would quantify how much judge error is affecting your eval scores and whether the hard tool-calling scenarios are being scored reliably.

In practice: A research-grade contribution with immediate practical relevance for anyone running LLM-as-judge evals. The multi-difficulty framing is the key insight: single-difficulty benchmarks have been hiding judge quality problems for a long time.

For: Both runtimes. Language-agnostic evaluation framework; the benchmark code and rubrics can be used to test any agent system including Axion and Hermes agents.

Security4
Quality3
Auditability3
Useful to you3
Useful to community4
Buildable now2
Hermes3

Verdict: watch. Score 22. Research paper with no library release yet. Watch for a pip-installable evaluation package. Once available, run it against the Axion scout eval judge to quantify baseline judge error before building on top of it.

Build #16 AgentJudgeBench: use the ai-implementation-build-intake skill to build this safely. Source: https://arxiv.org/abs/2608.26623. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · AgentJudgeBench submitted to HuggingFace Daily Papers on 2026-09-02 as paper 2608.26623; from ServiceNow-AI; benchmark tests LLM judges across easy, medium, and hard agentic tool-calling difficulty levels; 1 community discussion on HuggingFace papers page.

#17
watchBoth runtimes22 / 35

LangGraph SDK v0.4.4 repo

LangGraph SDK version 0.4.4, current Latest release. This is a patch release on the SDK component of the LangGraph framework; the release page contains minimal notes, indicating a maintenance or compatibility update rather than a feature addition.

What it does for you: LangGraph is not currently in the Axion stack, but it is the most widely used multi-agent orchestration framework for complex directed graph workflows. This version is worth monitoring if you evaluate LangGraph as an alternative to AutoGen for the Axion brain ingest pipeline. The SDK-only release scope means no breaking changes are expected.

In practice: A quiet maintenance release. Worth noting for version tracking but not actionable on its own.

For: Both runtimes. LangChain's agent orchestration SDK; provider-agnostic and compatible with Claude and Hermes workflows.

Security3
Quality3
Auditability3
Useful to you3
Useful to community4
Buildable now3
Hermes3

Verdict: watch. Score 22. Maintenance release with no visible feature additions. Monitor for the next feature-bearing LangGraph release and evaluate at that point.

Build #17 LangGraph SDK v0.4.4: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/langchain-ai/langgraph/releases/tag/sdk%3D%3D0.4.4. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · langgraph-sdk==0.4.4 tagged as Latest on langchain-ai/langgraph as of 2026-09-03; SDK-component-only release with minimal release notes; LangGraph main repo has 13,000+ stars.

#18
watchClaude (Axion)22 / 35

claude-code-setup plugin

Official Anthropic plugin for setting up new Claude Code projects. Provides scaffolding for CLAUDE.md, settings.json, hooks, and project structure following Anthropic's recommended conventions.

What it does for you: Useful when creating a new Claude Code project from scratch. The AXION project is already established, so this plugin is most relevant for any new isolated project or tool that spins up its own Claude Code context. Not present in YY's enabledPlugins as of 2026-09-03.

In practice: Low urgency for existing projects but immediately valuable for any greenfield Claude Code context. The TOOL-AND-APP folder sees new projects regularly.

For: Claude (Axion). Official Anthropic Claude Code plugin; Claude Code environment only.

Security4
Quality3
Auditability4
Useful to you3
Useful to community4
Buildable now2
Hermes2

Verdict: watch. Score 22. Not urgent for existing AXION work. Enable when starting a new Claude Code project to get the recommended scaffold and compare it against the existing AXION structure.

Install: add claude-code-setup to enabledPlugins in ~/.claude/settings.json when starting a new project. Review before enabling.

Source · claude-code-setup folder confirmed in anthropics/claude-plugins-official/plugins tree as of 2026-09-03 fetch; Apache-2.0 license on host repo; not present in YY's enabledPlugins as of 2026-09-03.

#19
newwatchBoth runtimes21 / 35

Control-Data Flow Separation for Multi-Agent Prompts repo

University of Waterloo research paper proposing a control-data flow separation approach for stable prompt optimization in multi-agent LLM systems. The core idea is that agent instructions (control flow) and task-specific content (data flow) should be kept in separate prompt regions to prevent optimization of one from destabilizing the other during automated prompt tuning.

What it does for you: The Axion orchestrator dispatches workers with prompts that mix task instructions and content in the same string. This paper provides a principled architecture for separating those concerns, which would make Axion worker prompts more stable during automated improvement cycles and reduce the risk of a prompt optimization run accidentally breaking the control logic. Directly applicable to any prompt engineering work in the Hermes pipeline.

In practice: A practical insight from a credible academic source. The separation of control and data flows is intuitive once stated and immediately applicable to any complex agent prompt without requiring a framework change.

For: Both runtimes. Framework-agnostic prompt engineering technique; applies to any multi-agent system including Axion orchestrator-worker patterns and Hermes pipelines.

Security4
Quality2
Auditability3
Useful to you3
Useful to community3
Buildable now2
Hermes4

Verdict: watch. Score 21. Research paper with no code release confirmed yet. Read the paper and apply the control-data separation principle to the next Axion orchestrator prompt you write or refactor.

Build #19 Control-Data Flow Separation: use the ai-implementation-build-intake skill to build this safely. Source: https://arxiv.org/abs/2609.00621. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · Paper 2609.00621 submitted to HuggingFace Daily Papers on 2026-09-02 from University of Waterloo; 6 upvotes on HuggingFace papers page; 1 community discussion; proposes control-data flow separation for stable prompt optimization in multi-agent LLM systems.

#20
newwatchBoth runtimes20 / 35

MCP Servers Python mcp-1.x requirement docs repo

A documentation commit to the modelcontextprotocol/servers repository, merged 2026-08-18, adding an explicit mcp 1.x version requirement note to the fetch, git, and time Python server READMEs. Previously undocumented; this clarifies that these servers require the mcp Python package at version 1.x and will not work with the 2.x line.

What it does for you: If any Axion MCP server is built against the reference fetch, git, or time servers from this repository, the version constraint is now explicit. Any pip environment that has upgraded to mcp 2.x would silently break these servers. The commit is now 16 days old; if it was relevant to your setup it would have surfaced already, but the mcp-server-dev plugin (rank 6) makes this worth noting.

In practice: A housekeeping documentation commit rather than a feature. Low urgency but worth knowing if you reference the official MCP server implementations.

For: Both runtimes. Documentation update to the official MCP reference server implementations; applies to any project building against the fetch, git, or time MCP servers.

Security4
Quality3
Auditability3
Useful to you3
Useful to community3
Buildable now2
Hermes2

Verdict: watch. Score 20. Documentation-only commit, 16 days old. Verify that any Axion MCP server built against the reference fetch, git, or time servers is pinned to mcp 1.x before enabling the mcp-server-dev plugin for a new build.

Build #20 MCP Servers Python mcp-1.x docs: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/modelcontextprotocol/servers/commit/d811ba11f1243a55bcc2e78d72ddd9b63b25f25e. Save canonical skill/agent under AXION\Skills and AXION\Agents.

Source · Commit d811ba1 merged 2026-08-18 by olaservo on modelcontextprotocol/servers; PR title: 'docs(python): note the mcp 1.x requirement in fetch, git, and time'; co-authored by claude; changes README documentation for 3 reference server implementations; mcp-below-2 cap enforced by companion commit 599dafc.