Docs, notes, workouts, journals in one place, searchable and cross-linked. A built-in agent can read the whole graph and propose edits you approve, or connect your own tools over MCP. Works fine without either.
Build an agent scoped to that one thing. It reads your graph, so it remembers what you logged (your workouts, your decisions, last month’s notes) and picks up where you left off every time. Ask it to remember, or to help you work out what it all adds up to.
Useful on its own. Better with agents on top.
An orbit-map of typed nodes (notes, journals, workouts, checklists, kanban boards, spreadsheets) with tags, cross-links, quick-capture, and full-text search. A working life-OS on its own, before you turn on any AI.
The built-in agent reads your whole graph and proposes edits you approve, with no copy-paste. Build your own, each scoped to one folder: a coach on your workouts, a companion on your journal. Or connect Claude and Cursor over MCP so your coding tools read the same graph.
It runs on your own OpenRouter key, and nothing you write is ever used for training. Every AI change waits for your approval, and you can export the whole graph whenever you want.
You capture on your phone and plan at your desk. Both touch one graph, so nothing you type tonight goes missing.
The gift idea hits while you wait to pay. Tap create, type nine words, done. It sits in your inbox checklist before the cashier scans your milk.
You open Push Day at the rack. Last week reads 135 for 5. You load 140, log 5 reps, and start your rest timer.
Your agent proposes splitting the auth module. You read the plan card, tap approve, and the edit lands before your stop.
Tap a card to descend, a bar to climb.
Your whole graph sits behind an authenticated endpoint, reachable and writable from any machine.
One graph behind an authenticated endpoint, reachable from Claude, Cursor, a script, or the web, on any machine.
Typed nodes, tags and edges, subtree and delta queries, per-key rate limits, and an audit trail, with full read and write access for agents.
Every tool and session shares the same graph, so context carries across machines and chats.
Wire your coding agent to the graph, then give it a standing rule. It makes one docs pass at each natural stopping point: task done, ready to commit.
One command connects the graph. A CLAUDE.md rule keeps it maintained.
claude mcp add --transport http aiyself \ https://aiyself.com/api/mcp \ --header "Authorization: Bearer <your-key>"
# CLAUDE.md Know where the fix goes? Go straight to code. Query the graph when you need context the code lacks: call flow, past decisions, gotchas. When a task wraps up or you are about to commit, make one docs pass: update the doc nodes for the files you touched (get_nodes fetches a batch in one call). Record flow, decisions, and gotchas.
To enforce it, add a Stop hook in .claude/settings.json that holds the session open until the doc node is updated.
Six moments from a normal week. Every edit waits for your approval.
You open Cursor on a project after two weeks away. Claude pulls your decision notes and open tasks over MCP before you type a line. You go straight to the fix.
Between sets you ask the coach what you benched last Tuesday. It reads your log, answers 135 for 5, and proposes 140 today as a card. Approve and load the plate.
Planning October in Lisbon, your travel agent opens the Kyoto folder first. It keeps the packing list that worked, drops the three things you never wore, and drafts the itinerary as a checklist.
Friday night you hit cmd-K: move everything I shipped this week to done. The agent proposes the card moves, you approve them inline, and all four boards are honest again.
The journal companion spots the word dread in four of your last five Sunday entries, each written the night before standup. You move prep to Friday and check back in a month.
The old chat starts forgetting your allergies, so you close it for good. Those details live as nodes now, tagged and searchable. Every new session starts already knowing you.
Paired Claude Opus agents ran identical coding tasks; one could read its project graph over MCP. The connected agent’s speed advantage climbs with how much context the task demands.
At the top of the curve the connected agent also answered three times as many tasks correctly: open decisions, ruled-out approaches, and deploy state live in the graph, so it reads them instead of rediscovering them.
From paired runs on identical tasks, July 2026.