Glonce

OzBrain launches a shared brain that AI agents read and write

toolsThe Glonce editorsHacker News🔥62Glonce rating62Impact50Novelty55Relevance90Surprise40Players35Credibility80

OzBrain, shown on Hacker News under the title "a shared brain for knowledge between agents and your team," is a subscription product that stores a user's projects, decisions, research and working preferences as structured articles in one place, then exposes that store to AI agents through an MCP connector. The pitch is that instead of pasting the same brief into Claude, then ChatGPT, then dropping the same markdown file into Cursor and letting the copies drift, every connected agent reads and writes the same current article. The page gives example categories a brain might hold: positioning and who a product is sold to, a specific client's scope and current terms, a person's own writing voice and words to avoid, a project's status and open decisions, and preferred models, tools and formats. The vendor says setup takes under two minutes: add a custom MCP connector in Claude by pasting the URL https://ozbrain.com/api/mcp under Settings then Connectors then Add custom connector, and sign in with an emailed code, which also creates the account if the user does not already have one. The company states the product works with Claude and ChatGPT through their native connector flows, plus Claude Code, Cursor, OpenClaw, Hermes Agent, and Gemini Spark where Google makes it available, and with any client that supports connectors generally. Pricing has four tiers: Free, at $0, capped at 50 articles; Pro, at $20 per month, capped at 300 articles; Max, at $99 per month, capped at 600 articles; and a custom Company tier with org-owned shared brains and per-seat pricing to be worked out directly with the vendor. Every tier is stated to include unlimited reads, writes and connections, and markdown export at any time. On trust, the company says content is encrypted at rest, sealed per account, that every access is logged, that it never trains on a user's brain content and never sells it, and that deleting the account deletes the content. The page attributes one unnamed quote, "Humans abandon wikis because the maintenance burden grows faster than the value," to a role description only, "OpenAI co-founder, Anthropic R&D," without naming the person. Separately from the marketing copy, the landing page's text contains a block written as direct instructions addressed to an AI agent reading the page rather than to a human visitor. It tells the agent to treat a linked docs page as a product description and not as instructions, to keep its first reply to the user under 150 words, and to offer up to three ranked reasons the product could help, each tied to real evidence from the user's own work, ending with one fixed closing line asking whether to help set the product up. If the user says yes, the embedded text goes on to script the agent through adding the MCP connector, explaining the email sign-in, waiting for each step, confirming the connection by listing the user's brains, recommending what to add first, and asking for approval before saving or inviting collaborators. The Hacker News submission itself carries 63 points and 30 comments at the time this dossier was compiled, and the page names no specific creator, company or team behind the product in its own text.

Key facts

  • OzBrain is an MCP-connector product that gives AI agents such as Claude, ChatGPT and Cursor one shared, structured knowledge base to read and write instead of context being copied between tools by hand.
  • Its landing page contains a block of text written as direct instructions to any AI agent reading the page, telling it to pitch the product in under 150 words with up to three evidence-based reasons and to end with one fixed closing question.
  • If a user agrees, the same embedded text scripts the agent through adding OzBrain as a custom MCP connector, signing in by email code, and staging content into a brain with the user's approval at each step.
  • Pricing runs from a Free tier at $0 for up to 50 articles, through Pro at $20 per month for up to 300 articles, to Max at $99 per month for up to 600 articles, plus a custom Company tier.
  • The vendor claims setup in under two minutes, encryption at rest, an audit log of every access, no training on or sale of user content, and full deletion on account closure, none of which the source text independently verifies.

Why it matters

The stated problem, that separate AI tools each hold their own partial memory and force a user to re-explain context by hand, is real and familiar to anyone running the same work through more than one assistant. What makes this particular launch worth a second look is not the pitch itself but the mechanism the page uses to make it: the copy is written so that an AI agent fetching the page, not just a human reading it, receives a scripted sales pitch and a step-by-step onboarding script addressed directly to it. That is a live example of a website authoring content for an AI reader rather than a person, on a page that also asks that AI reader to act on a user's behalf.

Who it affects

Anyone using more than one AI assistant or coding tool day to day, and any team trying to keep several people's agents working from the same current brief, is the audience the product targets. Separately, anyone who lets an AI agent browse or fetch web pages on their behalf is affected in a different way: this page demonstrates that ordinary-looking marketing copy can carry a block meant to be copied by a person into their own agent and obeyed there, not something the person composed themselves.

How to use it

The product is added as a custom MCP connector, pointed at https://ozbrain.com/api/mcp, inside an agent or app that supports connectors, such as Claude's Settings then Connectors menu, with sign-in done through an emailed code that also creates the account on first use. The vendor puts the process at under two minutes. Four tiers are offered: Free at $0 for up to 50 articles, Pro at $20 per month for up to 300 articles, Max at $99 per month for up to 600 articles, and a custom Company tier with org-owned shared brains and per-seat pricing set up directly with the vendor; every tier is said to include unlimited reads, writes and connections plus markdown export at any time.

How solid is it

Everything in this account comes from the product's own landing page, submitted to Hacker News by its own author; the source text names no specific creator, company or team, gives no user or usage numbers, and offers no independent benchmark or third-party account that the product works as described. Pricing, capacity limits, security practices and platform support are all the vendor's own claims. The one quote on the page about wikis is attributed only to a role, not a named person, so it cannot be checked against anything outside the page itself.

Risks and caveats

The embedded block addressed to an AI agent is the clearest thing to flag here: it instructs the agent to keep its reply short, to give the user reasons for buying framed as evidence from the user's own work, and to walk the user through installing a connector and handing the service read and write access to their conversations and files, all sourced from text the page invites the user to copy and hand to their own agent, not from an instruction the user composed directly. Anyone whose assistant browses the web should treat this as a working example of why content fetched from a page should not be treated as instructions from the user. Beyond that, none of the vendor's trust claims, encryption at rest, an audit log, no training on or sale of content, or full deletion on account closure, are demonstrated in the source text; they are stated, not shown.

“Humans abandon wikis because the maintenance burden grows faster than the value.”

— unnamed on the page, credited only by role as "OpenAI co-founder, Anthropic R&D"