Daily AI Implementation Scout Council

Today's ranked top 20

#1

build nowClaude (Axion)31 / 35

Claude Code September 2026 release train (v2.1.259 to v2.1.263) repo

Five point releases of Claude Code in six days, adding a skill doctor command, higher output limits, and a headless permission flag.

What it does for you: The new --permission-prompts none flag auto denies anything that would otherwise hang waiting for approval, which is exactly the failure mode an unattended cron job like the nightly brain ingest or the daily scout run would hit. The new /skill-doctor command tells you which of your 710 mapped skills are actually being loaded and used, so you can prune dead weight instead of guessing. bashOutputMaxChars and taskOutputMaxChars both rose to 128K, useful since Axion subagents sometimes truncate long tool output today.

In practice: Small, boring, immediately useful. The kind of release that fixes a problem you already have.

For: Claude (Axion). This is the Claude Code CLI itself, the harness Axion runs on.

Security5

Quality5

Auditability4

Useful to you5

Useful to community4

Buildable now5

Hermes3

Verdict: build now. Zero install cost, it ships with the next Claude Code auto update, and two of its features map directly onto known Axion pain points, headless hangs and skill bloat.

Build #1 Claude Code September 2026 release train (v2.1.259 to v2.1.263): use the ai-implementation-build-intake skill to build this safely. Source: https://code.claude.com/docs/en/changelog. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Version 2.1.263 shipped September 6, 2026, five days before this report, five point releases landed between September 1 and September 6.

#2

newtest firstClaude (Axion)26 / 35

Claude Developer Platform GA: Computer Use, Browser Use, Skills API, Files API repo

Anthropic moved Computer Use, Browser Use, the Skills API, and the Files API out of beta into general availability.

What it does for you: Skills API going GA means Agent Skills are now a stable, production grade API surface, not a preview, so building a skill driven Axion service directly on the API becomes safe to commit to instead of relying only on the CLI's SKILL.md convention. Files API GA means an agent can return a finished file straight through the API rather than YY manually collecting it from disk. Computer Use and Browser Use GA give a first party, billed, tool for an agent to drive a real screen or browser when a task needs it.

In practice: Strategically important, not something to flip on today without a billing and scope review first.

For: Claude (Axion). Native Anthropic API surface, only usable through Claude models.

Security4

Quality5

Auditability3

Useful to you4

Useful to community4

Buildable now3

Hermes3

Verdict: test first. Real capability unlock but each of the four APIs bills separately and GA sometimes changes rate limits versus the preview, worth a small test call before wiring into anything production.

Build #2 Claude Developer Platform GA: Computer Use, Browser Use, Skills API, Files API: use the ai-implementation-build-intake skill to build this safely. Source: https://www.channelinsider.com/ai/news-anthropic-claude-computer-use-skills-files-api/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · GA announced August 19 to 20, 2026, about three weeks before this report, coming soon to Vertex AI per the same announcement.

#3

test firstClaude (Axion)28 / 35

karanb192/claude-code-hooks plugin

An installable Claude Code plugin marketplace bundling 20 hooks for safety, cost, observability, and productivity.

What it does for you: One command, /plugin marketplace add karanb192/claude-code-hooks, gives access to a 20 hook catalog covering the same ground YY already hand builds hooks for, build-gate.js, ghl-write-gate.js, heartbeat.py. Worth mining for patterns or ready made guardrails rather than writing every hook from scratch, especially for categories Axion has not covered yet like cost tracking.

In practice: A genuinely useful reference catalog, but it is unaudited third party code running inside your Claude Code loop.

For: Claude (Axion). A Claude Code plugin marketplace of hooks, only runs inside Claude Code's hook system.

Security3

Quality4

Auditability4

Useful to you5

Useful to community4

Buildable now5

Hermes3

Verdict: test first. High fit and low install cost, but each hook runs arbitrary shell commands next to existing safety hooks like ghl-write-gate, read every hook before enabling it, do not bulk install.

Install: run /plugin marketplace add karanb192/claude-code-hooks in Claude Code, then review and enable specific hooks from the 20 plugin catalog one at a time. Review before enabling. Copy

Source · 367 GitHub stars, 20 plugin catalog, companion documentation site at hooks.karanbansal.in.

#4

test firstBoth runtimes25 / 35

Cognee v1.5.4 repo

An open source, self hosted memory platform that builds a knowledge graph plus vector index over ingested data.

What it does for you: Gives Axion a self hostable graph plus vector memory backend for cases where agents need to reason over relationships, client history, project docs, cross report links, rather than flat semantic search alone. Recent releases specifically hardened large dataset migration, relevant if the YY and MCL brain vaults keep growing the way they have been.

In practice: Promising but still stabilizing, the release cadence reads like an active rewrite, not a settled product.

For: Both runtimes. Self hosted Python service with a plain API, usable from Axion or Hermes.

Security3

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now3

Hermes4

Verdict: test first. Real fit for the brain vault use case, but self hosting a graph database is new infrastructure to maintain, worth a small pilot before committing.

Build #4 Cognee v1.5.4: 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 · Version 1.5.4 published to PyPI September 4, 2026, roughly 28,000 GitHub stars and 127 releases as of mid 2026.

#5

test firstBoth runtimes27 / 35

LangGraph 1.2.11 with native MCP support repo

A graph based orchestration library for stateful, multi step, multi agent LLM applications with durable execution and checkpointing.

What it does for you: Gives a state machine model for agent workflows with built in checkpointing, so a long running agent can pause, resume, or recover from failure without a hand rolled persistence layer. The new langchain.mcp namespace, built on FastMCP, means an Axion side experiment could call the same MCP servers YY already runs from inside a LangGraph flow instead of Claude Code's own orchestration.

In practice: Mature and capable, but heavier than anything Axion currently runs, a real framework switch, not a drop in add.

For: Both runtimes. A Python graph orchestration library, usable from any Python based runtime, not tied to Claude Code specifically.

Security4

Quality5

Auditability3

Useful to you3

Useful to community5

Buildable now3

Hermes4

Verdict: test first. Strong engineering and real new MCP support, but adopting it means buying into a second orchestration model alongside Claude Code, worth a small spike before any real build.

Build #5 LangGraph 1.2.11 with native MCP support: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/langchain-ai/langgraph. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Repo has 41.2k GitHub stars, current release is langgraph 1.2.11 with langgraph-sdk 0.4.4 released August 27, 2026, MCP support landed around September 1, 2026.

#6

watchBoth runtimes21 / 35

Microsoft Agent Framework python-1.17.0 repo

Microsoft's unified successor to Semantic Kernel and AutoGen, an SDK for building and orchestrating single and multi agent systems.

What it does for you: Gives agent orchestration, approvals, history replay, and hosting primitives with a real production support story from Microsoft. Only relevant to YY as a competitive reference right now since Axion is Claude Code first, not something to adopt without leaving the current harness.

In practice: Solid engineering, wrong ecosystem for where Axion lives today.

For: Both runtimes. Python and .NET SDK, framework agnostic, but a full alternative to Axion's current harness rather than something that plugs into it.

Security4

Quality4

Auditability3

Useful to you2

Useful to community4

Buildable now2

Hermes2

Verdict: watch. Well built and actively shipping, but low fit for a Claude Code centered stack and this release contains a labeled breaking change to migrate around.

Build #6 Microsoft Agent Framework python-1.17.0: 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 · Release python-1.17.0 shipped September 3, 2026, repo has 13.4k GitHub stars, reached GA 1.0 on April 3, 2026.

#7

newwatchClaude (Axion)21 / 35

Anthropic Memory Tool GA and unified chat plus Cowork memory repo

A server managed memory tool that lets Claude persist notes across sessions, now generally available, plus a merged chat and Cowork memory shipped for consumer plans.

What it does for you: This is the memory system living inside the same platform Axion is built on, so any future Axion agent built on the Claude Agent SDK could adopt the same pattern without a third party dependency. Directly usable today, however, only for Claude chat and Cowork, Anthropic explicitly excludes Claude Code from the unified memory rollout, so it does not touch Axion's coding agent workflows yet.

In practice: Worth knowing, not yet actionable for the coding agent side of Axion.

For: Claude (Axion). Native Claude API and consumer product feature.

Security4

Quality5

Auditability3

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. Explicitly excludes Claude Code, the surface YY actually runs Axion on, revisit once the SDK exposes the same memory tool for coding agents.

Build #7 Anthropic Memory Tool GA and unified chat plus Cowork memory: use the ai-implementation-build-intake skill to build this safely. Source: https://platform.claude.com/docs/en/agents-and-tools/tool-use/memory-tool. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Unified chat and Cowork memory shipped August 25, 2026, on by default for Free, Pro, and Max plans, memory tool works on all Claude 4 and later models.

#8

test firstBoth runtimes27 / 35

Langfuse v4 repo

An open source LLM engineering platform for tracing, evals, and prompt management, just shipped a ground up backend rewrite.

What it does for you: Becomes the dashboard that answers why an agent run cost so much or which prompt version regressed, without hand grepping runs.log. The new monitoring alerts on cost, quality, and latency pushed to Slack, webhooks, or GitHub Actions could sit next to the existing heartbeat.py failure alert pattern for the bridge.

In practice: Real engineering step forward, worth a self hosted pilot before wiring into anything live.

For: Both runtimes. Self hostable open source platform, works with any LLM runtime that can emit traces.

Security4

Quality5

Auditability3

Useful to you4

Useful to community4

Buildable now3

Hermes4

Verdict: test first. Directly useful for debugging the scout pipeline and the bridge, but v4 is a from scratch rewrite, worth testing before trusting it on anything production.

Build #8 Langfuse v4: use the ai-implementation-build-intake skill to build this safely. Source: https://langfuse.com/changelog/2026-08-17-langfuse-v4. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Version 4.0.0 reached general availability August 17, 2026, repo sits at roughly 34,300 GitHub stars as of September 7, 2026.

#9

newwatchBoth runtimes23 / 35

Promptfoo, now under OpenAI repo

The open source prompt and agent eval, red team, and regression testing tool that OpenAI acquired in March 2026 and has kept open source.

What it does for you: This is the tool behind Axion's existing prompt-eval-promptfoo skill, so the news is mainly an ownership change to track, not a new capability to adopt. Continued active releases since the acquisition mean the tool keeps improving under new ownership.

In practice: No action needed today, worth a note that ownership changed under a skill YY already relies on.

For: Both runtimes. CLI and library, model provider agnostic, already has an Axion skill built on it.

Security3

Quality4

Auditability3

Useful to you3

Useful to community4

Buildable now3

Hermes3

Verdict: watch. Already adopted through an existing skill, the only new information is the OpenAI acquisition and the roadmap risk that comes with it.

Build #9 Promptfoo, now under OpenAI: use the ai-implementation-build-intake skill to build this safely. Source: https://www.promptfoo.dev/blog/promptfoo-joining-openai/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Acquisition announced March 9, 2026, latest npm version 0.122.2 published August 28, 2026, prior to acquisition it had over 350,000 developers and over 25 percent of the Fortune 500 as customers.

#10

newwatchBoth runtimes19 / 35

NeMo Guardrails on NVIDIA-NeMo repo

A programmable guardrails toolkit for LLM systems covering input, output, dialog, and tool use rails through a DSL called Colang.

What it does for you: Would sit in front of an Axion agent to deterministically block unsafe tool calls or off topic output instead of relying only on a system prompt. Reports of a 40 percent runtime overhead cut versus the prior version matter if rails are checked on every agent turn, but Axion has no rails layer today, so this would be new infrastructure, not a swap in.

In practice: Solid tool, no current gap it fills for Axion, the existing hook based gates cover similar ground already.

For: Both runtimes. Python toolkit with its own Colang DSL, runtime agnostic but requires its own rails engine.

Security4

Quality4

Auditability2

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. Real engineering improvement but Axion's guardrail needs are currently met by purpose built hooks like ghl-write-gate, not a general rails DSL.

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

Source · Latest tagged release v0.24.0, project moved to the NVIDIA-NeMo GitHub organization from the older nvidia and nemo-guardrails path, Apache 2.0 licensed.

#11

newtest firstClaude (Axion)25 / 35

Cloudflare MCP servers plus Skills plugin on spec 2026-07-28 plugin

Cloudflare's official remote MCP servers, rebuilt on the new stateless MCP spec, bundled into a Cloudflare Skills plugin with slash commands for Claude Code.

What it does for you: YY already runs the AI scout newsletter on Cloudflare Pages, one OAuth connection at mcp.cloudflare.com/mcp would give Claude Code direct control over that account, DNS, Workers, KV, instead of the current manual deploy step with a token in cloudflare.env. The bundled skills mean Claude knows how to use each domain server correctly out of the box.

In practice: Genuinely convenient for an account YY already has, but broad account access needs scoping first.

For: Claude (Axion). Bundled as a Claude Code Skills plugin with slash commands.

Security3

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Direct fit for the existing Cloudflare Pages deploy, but OAuth grants broad account access, scope it to a least privilege API token before connecting.

Install: in Claude Code run /plugin marketplace add cloudflare/deploy-mcp, or add https://mcp.cloudflare.com/mcp as a remote MCP server and complete the OAuth prompt with a scoped token. Review before enabling. Copy

Source · Remote endpoint at mcp.cloudflare.com/mcp supports MCP spec 2026-07-28 with OAuth based auth, upgraded to the new stateless transport and a Code Mode option this cycle.

#12

newwatchClaude (Axion)18 / 35

Claude Commerce Agents blueprint plugin

Anthropic's official blueprint with a shopping agent and a merchant agent, shipped with a Claude Code plugin to scaffold it against real systems.

What it does for you: A ready made pattern for a customer facing shopping agent that searches and builds carts, plus a staff facing merchant agent that drafts listings and pricing for human approval, across retail, travel, telecom, and ticketing verticals. Neither MCL's coaching business nor YY's personal tools currently have a checkout or catalog flow, so there is no live use case today.

In practice: Well built reference architecture, no current fit.

For: Claude (Axion). Ships as a Claude Code plugin with a bundled scaffolding installer.

Security3

Quality4

Auditability3

Useful to you1

Useful to community3

Buildable now2

Hermes2

Verdict: watch. High quality Anthropic reference blueprint, but MCL's business model has no commerce or checkout flow for it to plug into right now.

Install: clone github.com/anthropics/commerce-agents, then run its bundled Claude Code plugin installer per the repo README. Review before enabling. Copy

Source · 2.2k GitHub stars, released September 2, 2026, Apache 2.0 license, requires Python 3.11 or later and Node 22.

#13

watchBoth runtimes20 / 35

MCP Specification 2026-07-28, stateless core repo

The current official MCP protocol spec revision, moving the protocol core from stateful sessions to a stateless request and response model.

What it does for you: This is the standard every MCP server Axion talks to eventually runs on. The stateless core removes the session affinity problem that broke autoscaled MCP servers, and it starts a 12 month deprecation clock on Roots, Sampling, Logging, and Dynamic Client Registration, worth an inventory check against any custom MCP server Axion runs.

In practice: Foundational and worth tracking, nothing to build directly from it today.

For: Both runtimes. A protocol specification, language and runtime agnostic by design.

Security4

Quality5

Auditability2

Useful to you2

Useful to community4

Buildable now1

Hermes2

Verdict: watch. Informational and foundational rather than something to build, revisit once Axion's own MCP servers or dependencies need to migrate off the deprecated features.

Build #13 MCP Specification 2026-07-28, stateless core: use the ai-implementation-build-intake skill to build this safely. Source: https://blog.modelcontextprotocol.io/posts/2026-07-28/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Released July 28, 2026, with a formal 12 month minimum deprecation window before the older stateful transport and features are removed.

#14

newwatchBoth runtimes20 / 35

Deadbugz MCP supply chain campaign repo

An active malicious MCP server distributed via GitHub pull requests, disguised as a harmless productivity suite server.

What it does for you: Nothing to install, this is a warning directly relevant to Axion because it targets exactly the kind of Claude Code plus MCP setup YY runs. It behaves normally for the first three tool calls, then rewrites its own tool metadata to steer the agent toward reading local files while hiding the activity.

In practice: Genuinely unsettling, a strong argument for never installing an MCP server from an unverified pull request.

For: Both runtimes. A threat pattern relevant to any MCP consuming runtime, not a tool tied to one platform.

Security5

Quality4

Auditability3

Useful to you3

Useful to community3

Buildable now1

Hermes1

Verdict: watch. Not a build candidate, filed as an active security signal to keep in mind whenever a new MCP server is considered for Axion.

Build #14 Deadbugz MCP supply chain campaign: use the ai-implementation-build-intake skill to build this safely. Source: https://www.pillar.security/blog/deadbugz-currently-active-mcp-supply-chain-campaign. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Pillar Security tracked 23 malicious pull requests filed in a 74 minute window on August 10, 2026, 17 remote MCP config submissions, 4 local script configs, 2 directory listing submissions.

#15

newwatchBoth runtimes21 / 35

A2A Protocol under the Agentic AI Foundation repo

An open protocol, now stewarded by the Agentic AI Foundation, that lets agents built on different frameworks discover each other and hand off tasks.

What it does for you: If Axion ever needs to talk to a third party agent it does not control, a vendor's support agent or a partner's ordering agent, A2A gives a standard agent card plus task handshake instead of a bespoke API per partner. No such integration exists today, so this is a watch item, not a build.

In practice: The closest thing to a real cross vendor agent standard right now, worth revisiting once a concrete partner integration comes up.

For: Both runtimes. An open protocol with multi language SDKs, not tied to any single runtime.

Security3

Quality4

Auditability3

Useful to you2

Useful to community4

Buildable now2

Hermes3

Verdict: watch. Real standard with active SDK development, but no current Axion use case needs cross vendor agent handoff.

Build #15 A2A Protocol under the Agentic AI Foundation: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/a2aproject/A2A. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Repo sits at 25.6k GitHub stars as of September 2, 2026, sub repos for dotnet, js, and java all show commits in the September 4 to 6, 2026 window.

#16

newwatchBoth runtimes15 / 35

Agent Payments Protocol AP2, FIDO Alliance governance transfer repo

Google and Coinbase's open protocol for AI agents to make verifiable, auditable payments, now transferring governance to the FIDO Alliance.

What it does for you: This would matter if an Axion agent were ever meant to actually buy something on YY's behalf with a paper trail, but no such payment flow exists in Axion or MCL's business today. Filed for awareness that a real standards body is now behind it.

In practice: Well governed and maturing, no current need for it.

For: Both runtimes. A protocol spec with reference implementations in multiple languages.

Security3

Quality4

Auditability2

Useful to you1

Useful to community3

Buildable now1

Hermes1

Verdict: watch. No payment agent use case exists today, revisit if Axion or MCL ever need an agent to transact on their behalf.

Build #16 Agent Payments Protocol AP2, FIDO Alliance governance transfer: use the ai-implementation-build-intake skill to build this safely. Source: https://ap2-protocol.org/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Version 0.2.0 shipped April 2026 with Python, TypeScript, Kotlin, and Go reference implementations, over 60 partner organizations reported integrating including PayPal, Mastercard, Shopify, and Etsy.

#17

watchStandalone tool20 / 35

RAGFlow agentic RAG update repo

An open source RAG engine fusing document parsing, chunking, and retrieval with agent capabilities in one deployable stack.

What it does for you: Could be self hosted as an internal knowledge engine over coaching transcripts and SOPs, skipping a from scratch build of chunking plus retrieval plus UI. The August update added agentic RAG with four reasoning depth modes, trading cost for thoroughness per query.

In practice: Capable but heavy, and a data residency question worth resolving before it touches client files.

For: Standalone tool. A full self hosted product stack you deploy and use, not a library to integrate into Axion's own loop.

Security2

Quality4

Auditability2

Useful to you3

Useful to community4

Buildable now2

Hermes3

Verdict: watch. Real capability but a heavy Docker based self hosted stack with a Chinese origin data residency question, worth a closer look before any pilot with client material.

Build #17 RAGFlow agentic RAG update: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/infiniflow/ragflow. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · 81.8k to 90k GitHub stars depending on source, agentic RAG with four thinking modes shipped August 19, 2026, connector additions followed August 28, 2026.

#18

newwatchBoth runtimes21 / 35

jina-reranker-v3.5 repo

A 0.6 billion parameter listwise reranker that scores an entire candidate list in one forward pass instead of one document at a time.

What it does for you: Could be a drop in reranking step after any existing vector search, for example over the 2nd Brain vaults, meaningfully faster than the prior version and better on long documents up to 131,000 tokens. The non commercial license is a hard blocker for any paid MCL product, so this is personal use only unless a commercial license is purchased.

In practice: Technically strong, licensing rules out the business use case.

For: Both runtimes. A model usable via API or self host from any runtime.

Security4

Quality4

Auditability3

Useful to you2

Useful to community3

Buildable now2

Hermes3

Verdict: watch. Non commercial license blocks use inside anything MCL charges for, and Axion has no current reranking step it would slot into.

Build #18 jina-reranker-v3.5: use the ai-implementation-build-intake skill to build this safely. Source: https://jina.ai/en-US/news/jina-reranker-v3-5-faster-listwise-reranking-hybrid-attention-self-distillation/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Released July 27, 2026, scores 63.20 NDCG@10 on BEIR and 74.11 on MIRACL, runs 1.22x to 1.56x faster than the prior version.

#19

test firstClaude (Axion)26 / 35

agent-browser WebMCP support (Vercel Labs) repo

A CLI that drives Chrome via the Chrome DevTools Protocol using accessibility tree snapshots and compact element references, built for AI coding agents.

What it does for you: Gives Axion a lightweight, token efficient way to let Claude drive a real browser from the CLI for web QA or scraping, without wiring up full Playwright or a browser MCP server. It ships with a ready made Claude Code skill, so it could slot straight into a render-verify style check. The new WebMCP support lets the browser discover and call tools a target site exposes natively, cutting down on manual selector work.

In practice: Practical and immediately testable for Axion's existing render-verify habit.

For: Claude (Axion). Ships a bundled Claude Code skill and is built specifically for AI coding agents.

Security3

Quality4

Auditability4

Useful to you4

Useful to community4

Buildable now4

Hermes3

Verdict: test first. Ready made skill and strong npm install numbers, but WebMCP itself is experimental and depends on sites implementing it, worth a small pilot on a real render-verify task before relying on it.

Build #19 agent-browser WebMCP support (Vercel Labs): use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/vercel-labs/agent-browser. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · About 22.6k to 23.4k GitHub stars, 109,000 to 164,000 weekly npm installs, WebMCP support shipped September 1, 2026.

#20

newwatchBoth runtimes18 / 35

Supabase agent MCP write access plugin

Supabase's MCP server now supports full read and write access for agents, not just read only, turning a Postgres project into a state store an agent can act on directly.

What it does for you: This is the connector pattern that matters if a future Axion tool needs a durable, queryable state store instead of flat files or a vector store, an agent can read, insert, update, and run migrations against a real database mid conversation. No Supabase project exists in Axion today, so this is a watch item until a concrete need appears.

In practice: Powerful and exactly the kind of blast radius that already justifies YY's GHL two confirmation rule elsewhere.

For: Both runtimes. An MCP server, protocol based and usable from any MCP capable runtime.

Security2

Quality4

Auditability3

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. Full write access from an agent to a production database needs a confirm twice style gate before it touches real data, and no current Axion tool needs this connector yet.

Install: add Supabase's MCP server to a Claude Code MCP config and connect it to a project, full read and write access is granted by default so scope a service role key first. Review before enabling. Copy

Source · Became a live connector on Perplexity Computer on August 7, 2026, full create, read, update, delete, plus migrations confirmed via its MCP server.