What the connector can do
Each MCP tool, what it can read and write, and what is sent to OpenKnowledge Cloud.
What it can reach
A connected AI tool acts as the person who signed in, with that person's current permissions.
- It reaches only the brains that person can open in OpenKnowledge Cloud. A Viewer can read; changing files needs Editor or Owner access.
- It cannot reach anything on your computer. The
exectool runs shell commands against files stored in OpenKnowledge Cloud, not on your machine. It has no network access, andpython3,nodeandtarare not available. - Permission changes apply at once. When someone removes you from a brain, your AI tools lose it too.
Tools
The account-wide MCP address offers the tools below, among others. Reads only means the tool never changes anything, and the tool says so to your AI client with the MCP readOnlyHint annotation.
| Tool | What it does | Reads only |
|---|---|---|
projects | Lists the brains you can reach. Can also create a brain, rename one, and change its settings. | No |
exec | Runs a shell command (cat, ls, grep, sed, mv and similar) against a brain's files to read and edit them. | No |
search | Full-text search across a brain's pages and files. | Yes |
links | Lists links between pages: backlinks, outgoing links, broken links, and pages with no links. | Yes |
properties | Changes a page's frontmatter properties, such as tags, without rewriting the page. | No |
versions | Lists earlier versions of a page or uploaded file, and restores one. | No |
comments | Lists, starts, replies to, resolves and deletes comment threads. | No |
upload | Adds a file, such as an image or PDF, to a brain. | No |
read_upload | Gets an uploaded file's details and a short-lived download link. | Yes |
skills | Lists the skills saved in a brain, changes their properties, and imports a skill from skills.sh or GitHub. | No |
skill_registry | Searches public skills on skills.sh and GitHub before you import one. | Yes |
members | Lists who can access a brain, and adds, changes or removes their role. | No |
invitations | Invites people to a workspace by email and manages pending invitations. | No |
share | Gives an existing account Viewer access to one document, and removes it. | No |
shared_with_me | Reads or edits one document someone shared with you. | No |
Tools that can delete or overwrite data, such as exec, are marked with the MCP destructiveHint annotation, so your AI client can ask you before running them. skills, skill_registry and invitations are also marked openWorldHint, because they reach skills.sh, GitHub or send email.
A brain's own MCP address, from its settings, offers the same project tools for that brain only. It has no projects, members, invitations, share or shared_with_me.
What is sent to OpenKnowledge Cloud
When your AI tool calls a tool, it sends the tool's name and arguments to app.openknowledge.com. Arguments can include shell commands, the text of pages being written, and files being uploaded. The request carries your sign-in token and the name your AI client reports about itself.
OpenKnowledge Cloud keeps:
- The changes you asked for. Edits are saved to the brain with version history, like edits made in the app.
- Who made each change. Each write is credited to you and to the agent, named after your AI client, and appears in the page's timeline. A short
summarythe agent passes becomes the timeline entry. - Service logs. Each call logs the tool name, whether it succeeded, and an error code if it failed. The tool's arguments are not logged.
The tool's answer goes back to your AI tool, which handles it under its own provider's terms.