recall and explain are read-only, forget is marked destructive.12claude mcp add --transport http kaibo $KAIBO_URL/mcp \ --header "Authorization: Bearer $KAIBO_KEY"
/mcp shows whether kaibo connected. Or run kaibo init in your project and it writes a project .mcp.json for you.recall and answer July 17 with the source. If it does not see the tools, restart the agent, and check the key with kaibo doctor.X-Project-Id: <project id or slug> header to the server entry. The slug is the project's name in lowercase with dashes (try-new); kaibo projects shows both (Claude Code: --header "X-Project-Id: <id>", Cursor: add it under headers).kaibo mcp): set KAIBO_PROJECT=<project id> in the server's env.kaibo keys create --for-project <id> ...), and the dashboard's project settings page shows snippets with the project already filled in.read_write can remember and recall. Add forget only if the agent should delete. Never use an admin key in an agent./mcp. Send your key as a Bearer token. The key is passed straight through to the API and is never stored; scopes are enforced by the API, so a read-only key cannot forget.kaibo mcp over stdio and reads KAIBO_URL and KAIBO_KEY (and KAIBO_PROJECT for an organization-wide key) from its environment. Use it for tools that only start local servers.| Tool | Input |
remember | text, wait (default true) |
recall | query, asOf?, budgetTokens?, types? |
explain | memoryId, depth? |
forget | memoryId or query, reason, dryRun? |
recall text is wrapped between markers that include a random value generated per response, and the block is introduced as stored data. That makes it harder for stored text to fake the end of the block. It lowers prompt-injection risk and does not remove it. Treat recalled text as untrusted in your own agent.