OzBrain Puts a Review Gate Between AI Agents and Shared Knowledge

A team using several AI agents can now give them one shared collection of work context—and put a check between an agent’s proposed update and the record everyone else will read. OzBrain is a hosted knowledge base that connector-compatible agents can read and write. Its defining constraint is that new material is staged, while a conflict with an existing article pauses the write instead of changing the accepted version.

A shared record, rather than scattered memory

OzBrain organizes the knowledge base as linked articles, with a routing index intended to help an agent open material relevant to its task rather than load the whole library. Its examples include positioning, client information, writing voice, project status and tool preferences. The company presents that structure as a shared source of truth across agents, rather than a separate memory store inside each chat product.

The important mechanism is the write path. OzBrain says it routes a contribution to the relevant article and stages it before promotion. If the contribution contradicts canonical content, the service surfaces the conflict and leaves the existing record unchanged. Its example contrasts a proposed $49 monthly price with a canonical $29 price. Version history records the agent responsible for each version and the time of the change.

That design makes a disagreement over a client term or operating decision a review task, not an invisible overwrite. OzBrain’s setup instructions tell users to inspect a staged draft, make corrections if needed and approve it before promotion. Users or authorized agents can review staged changes.

The connector is the common access point

Compatible clients reach OzBrain through its custom MCP endpoint, a connector address that lets an agent access the same brain. OzBrain says the service works with Claude, ChatGPT, Claude Code, Cursor, OpenClaw, Hermes Agent and Gemini Spark where available, along with other connector-supporting clients. Claude uses a custom connector flow, while Cursor can add OzBrain as an MCP server.

Controls OzBrain says it provides

• Encrypted article bodies at rest, tenant isolation through forced Postgres row-level security, and revocation for connected agents.
• An audit log that records reads and writes and can be exported as CSV.
• Markdown export and hard deletion of account content, alongside a statement that customer brains are not used to train foundation models or sold for advertising.

Those controls apply to OzBrain’s service. Content sent onward to a connected AI service remains subject to that provider’s terms, a material boundary for teams considering shared client details, pricing decisions or internal project status.

50 articlesFree

The Free plan is listed at up to 50 articles.

300 articlesPro

Pro is listed at $20 per month for up to 300 articles.

600 articlesMax

Max is listed at $99 per month for up to 600 articles.

The bill tracks the library, not agent traffic

OzBrain describes reads, writes, brains and connections as unlimited on its Free, Pro and Max plans, subject to each tier’s article ceiling. The stated billing constraint is therefore the size of the maintained library, not the number of agent actions. A custom Company tier is listed for organization-owned shared brains above the Max limit, with per-seat pricing to be designed with customers.

The test is whether review can keep up

The product’s value depends less on whether an agent can retrieve a note than on whether a shared library remains useful as multiple agents add competing updates. OzBrain says it can flag stale material and refactor oversized articles into smaller ones. Its staged-write system supplies a visible control point, but teams will still need to decide which agents can propose or approve changes and how much review their shared knowledge warrants.