Agentic AI
Hub

What you're looking at. When Claude Code wants to do anything outside of reading and writing local files — create a Notion page, trigger a Make scenario, query the vault, create an Asana task — it doesn't speak directly to those upstream APIs. It calls a named MCP tool, and an MCP server's router takes over: pick the right handler, validate the input, hit the upstream system, normalize the response back into MCP's structured content format. Every tool you see in the demo is a real handler I can call from an agent right now.

Why a router instead of one endpoint per tool. 21 tools as 21 separate processes would mean 21 connections to maintain, 21 auth setups, 21 lifecycle bugs. A single server with grouped routing keeps the surface coherent — one process per integration domain — shares auth and rate-limit logic, and lets the agent discover related tools through MCP's capability negotiation instead of guessing names. When I add a new tool, the router picks it up automatically; nothing on the client side has to change.

Why Zod sits in the middle. The schema isn't just for nice TypeScript types — it's the runtime gate. Every input from the LLM is parsed against the schema before the handler runs. Wrong type, missing field, out-of-range number — caught here, not three API calls deep. The same schema is exported as JSON Schema to the client, so the LLM sees exactly what the server expects. One source of truth, two consumers (the runtime and the model).

Why pipelines, not just MCP tools. MCP tools are great for things Claude does while I'm actively working with it. Pipelines are for things that happen withoutme — a webhook fires at 3am, an event arrives while I'm asleep, a meeting ends and a transcript needs to land somewhere. n8n runs on a VPS I control, listens for those events, and does the work in the background. By the time I open my laptop, the work is done.

What each pipeline actually does. Idea Catcher turns voice memos from an iOS shortcut into structured Notion pages — Whisper transcribes, GPT-4o-mini cleans and tags, n8n routes to the right database, and a human-in-the-loop gate decides whether the idea becomes an Asana task. Lead Pipelinepolls a Notion booking-lead DB every minute, scrapes the lead's site, generates a discovery-call brief in Notion, and waits for me to review — the final “link into pipeline” step is still on the TODO list. YouTube Pipeline takes a topic from the vault, drafts a script with Claude, waits for me to approve, renders with Remotion, then waits again before upload — two gates, because publishing is irreversible.

Why the gates matter. Full automation sounds great until the AI mistakes a voice memo for a confidential note and posts it publicly, or the script generator hallucinates a fact about a real client. Every gate is a deliberate stopping point where a human checks before irreversible work happens. I keep automation honest by drawing clear lines between “AI can do this alone” and “AI proposes, human approves.”

What the agent actually does with this. When I ask Claude something like “why did we pick n8n over Make for the lead pipeline?”, it doesn't guess. It calls vault_query with the question, gets back a ranked list of relevant notes — decisions, learnings, project entries — and then answers, grounded in what I actually decided. No hallucinated history. No invented rationale. The graph you see is the result that comes back: nodes lit by relevance, connections showing related context.

Why semantic, not keyword. A keyword search misses everything where I phrased it differently — “automation platform” vs. “workflow tool”, “decision” vs. “trade-off”. Embedding-based semantic search matches by meaning, so the agent finds the relevant note even when the query and the title don't share a single word. That's the difference between an assistant that searches your notes and one that actually understands them.

The structure makes it work. Every note is atomic — one decision per file, one learning per file, one skill per file — with typed frontmatter (type, tags, project, created). The graph isn't a cosmetic visualization; it's the actual shape of the vault. When the agent searches, it can filter by note type before even looking at content (“only show me decisions about agents in the last quarter”), which is impossible against a wall of free-form Markdown.

What “agentic” means here. Most people's mental model of AI is a chat box: you ask, it answers. An agentic workflow is fundamentally different — you give it a goal (“deploy the new feature to staging”), and it figures out the steps: read the deploy checklist from the vault, run typecheck and tests in parallel, edit the staging config, push the branch, report the result. No further input from me until something actually requires a decision.

Why sub-agents matter. A single agent doing everything sequentially is slow and context-greedy. By spawning sub-agents — one for build validation, one for tests, one for code-search — the main agent can run independent work in parallel and keep its own context clean. Each sub-agent gets a narrow job and a small context window, which is faster, cheaper, and more reliable than one giant context full of unrelated tool calls. The orchestration view above shows exactly that: bars on parallel lanes, sub-agents indented under the main thread.

What this replaces. Without orchestration, every multi-step job needs me in the loop the whole time — paste output, tell it the next step, paste again. With orchestration, the multi-step job becomes a single instruction. The three workflows in the demo (Deploy, Research, Bug Fix) are real shapes I run regularly; what used to be 30 minutes of copy-paste between Claude and the terminal is now one prompt and a finished result.

Why this is built on MCP. Every action in the timeline — vault queries, file reads, edits, bash commands, git operations, sub-agent spawns — is an MCP tool call. The agent doesn't need a custom framework to orchestrate; it needs a typed surface of tools and the freedom to call them in whatever sequence the goal demands. MCP is the substrate; orchestration is what you get for free once the substrate is right.