AI clients
Connect Claude.ai, ChatGPT, Cursor, or Claude CLI from Settings, sign in safely, and review what each client can read or propose.
Use AI clients in Settings to let external assistants read your workspace through Knownwise’s MCP server, with the same guardrails as the in-app product.
Start on the Settings page
Open Settings and the AI clients section. The Connect an AI client card shows the Server URL, a Copy control, and short instructions for Claude.ai Connectors (URL only, not JSON).
Under Where to paste it, pick Claude.ai, Claude CLI, Cursor, or ChatGPT. Each tab lists numbered steps and a short note about that client. After you allow access, the client can read every brand in this workspace. It cannot publish, bill, or talk to the in-app agent. Read the MCP docs opens the MCP server reference for the full tool list.
Sign in from the client
When a client opens Knownwise to sign in, you land on Allow AI client access. You see what the client asked for under What it can do, including Read brand evidence: visibility, answers, sources and reports. You may optionally allow write proposals; they wait for your approval in Knownwise unless you let small changes apply automatically. Connected clients can see every brand in this workspace. Never shared lists what is excluded (agent chat memory, billing, buying capacity, merging pull requests, or publishing your site). Choose Allow access or Deny.
Claude CLI keys
Claude CLI has no URL dialog. In Claude CLI, choose Generate a key, enter a name, and decide whether to tick Allow changes. The box is off by default: without it, the key is read-only (Read only on the key list). With it, the client may propose new questions, competitors, and settings (Can propose changes); they wait for your approval in Knownwise, except small changes when automatic apply is on. Copy the one-time JSON and run the command shown for your machine. You can keep up to ten keys; Revoke removes one. Signed-in clients lists OAuth sign-ins with scopes and Revoke; Revoke a client and it loses access immediately.
Read tools return bounded brand evidence across your workspace; write tools propose changes that wait for your approval, apart from small changes when automatic apply is on. The full tool list lives in the MCP server reference.
Let small changes apply automatically
Every change an assistant proposes waits for your approval until you turn on Apply small changes without asking under AI clients. It is off by default, an assistant cannot turn it on, and it covers every client that may propose changes.
When it is on, three kinds of change apply at once: adding questions, resuming a question you stopped tracking, and setting a brand's Client group. They follow the dashboard's rules, so a question over your plan's active limit, one that names your brand or a competitor, or a duplicate is refused. Everything else still waits for your approval: editing a question or stopping it, adding a competitor, and any other brand setting. A request that mixes a small change with anything else waits as a whole.
Changes from assistants
Changes from assistants, under the switch, lists the newest 50 changes with the client, the brand, what each change does and its status. Review opens a change that is waiting for your approval. Undo takes back a change that applied automatically: the questions it added or resumed stop being tracked, and the client group goes back to what it was. If a question was edited or stopped, or the group changed, after the change applied, Undo refuses and changes nothing.