Second Brain
You have a knowledge base. Every time you open an AI chat, you become the API between it and your agents — copying the relevant notes in, by hand, again. Glassy closes that gap. Everything you save is indexed once, then queried by your AI tools directly, with citations.
A search engine answers questions. A second brain carries state: what was decided, what the agent was told, what it wants to do next. These four lanes are the difference, and all four are shipped.
Memory
An agent saves durable memories as real notes — searchable, tagged, and linked into your graph like anything you wrote yourself. Recall is scoped, so an agent gets back its own memory, not the whole workspace.
glassy_remember · glassy_recall
Canvas
Canvas is the page type that can hold live content. You and your agent arrange the same surface, and the embed policy that decides what may render is written down and reported — not hidden in a black box.
origin-allowlisted embeds, reported per surface
Approval
Every decision that needs a human lands in one queue, with the consequence of each choice written down. The note itself carries the verdict — approved, changes requested, under review — so you decide on the work, not only in a list.
review queue · 4 review tools (appliance)
Pairing
A calendar event can dispatch a task to a running agent. The result rides the event, appears in the activity feed, and can post itself back as a durable note — so the outcome is visible end to end.
event-triggered dispatch · activity feed
Same server, same tools. What changes is who holds the corpus and what the budget is.
Our cloud
Ships on every account, free included. One key per account, 37 tools, and a tool-call budget of 120/hour on Free and 1,200/hour on Pro and Clear.
Your corpus is indexed for you. Five tools stay appliance-only — the review queue and owner notifications need a machine you control.
Compare it →Your machine
One docker compose up. All 42 tools, a named key per agent so authorship and memory are attributable, and no tool-call budget at all.
Deep agentic loops run unthrottled because the only meter is your own hardware.
Compare it →Your own server
Link your appliance to your own instance and keep a server-to-server sync channel between them, with no Glassy server in the path.
Same capability as your machine, plus an instance that is yours to keep.
Compare it →Ask them in Claude, Cursor, Windsurf, OpenCode, Hermes, Openclaw, Pi, or any MCP client. Your assistant answers from your corpus, not from its training data.
Save
The Glassy Capture extension takes a page in one keystroke. Notes, documents and voice memos land in the workspace. Obsidian vault files sync in the background.
Index
Everything is embedded and indexed automatically. BM25 full-text search plus vector semantic search, merged with Reciprocal Rank Fusion.
Query
Your agent searches the corpus and returns real results with citations. 42 tools,
4 prompts and 9 resources on the official
@modelcontextprotocol/sdk v2.
The knowledge base
Sources, links, orphan counts, and the hubs your notes actually orbit — visible, countable, and yours to rebuild. Not a black box you have to trust.
No. Your corpus is never used as training data, and we do not sell it. On your own machine the corpus does not leave it at all — only the query you send goes to the assistant you chose.
No. Generate a key in Settings, copy the config snippet for your client, paste it in. The snippet carries the live server URL. That is the whole setup.
Second Brain works with Glassy-native bookmarks, notes and documents alone. Obsidian is an optional layer, not a requirement.