MCP connects tools to your workspace
MCP gives an agent a way to access external data and operations. In AURA, these are website projects, questions, reports and actions authorized by the owner. The protocol itself grants no access to computer files or another AI account. Reading your workspace uses AURA authorization. Website changes require the executor's separate capabilities. Distinguish three states: the server is added to the client, sign-in is completed, and the required tool actually succeeds for an authorized project. Only the last provides practical confirmation that the connection is ready.
A practical Codex example
The official Codex instructions describe remote HTTP connections and separate OAuth sign-in. For AURA, config.toml uses the [mcp_servers.aura] section with url = "https://aurageo.io/mcp". Then run codex mcp login aura and complete the authorization that opens. Do not put secrets in a public repository. After signing in, ask the agent to show available projects. If the list is empty, a project may not yet exist or may not have been selected during authorization; this is not a reason to assume data is lost and immediately create duplicates.
Other clients and the existing plugin
AURA documentation includes options for Claude Code, Cursor, Gemini CLI and other compatible tools. Their configuration formats differ. A client named Gemini does not automatically provide control of Gemini's web chat. A plugin may package an MCP connection and instructions conveniently, but does not remove sign-in, project permissions or the assistant's own capability requirements. Reuse an existing connection if it already works; changing the server identifier or creating a second plugin merely to read a project is generally unnecessary. Public distribution of a plugin is a separate owner action.
A readiness check that proves something
Ask the agent to read one accessible project and its onboarding, then briefly state the next step. Reading needs the relevant permission; saving an AI answer requires observe, while publication requires separate permissions and a verified website. A tool error should become a specific explanation: session expired, project not selected, operation unauthorized or limit reached. A general educational GEO article can be read without connecting a workspace. Connecting private projects to an agent is useful when the task requires your data or authorized changes.
Conditions and limitations
- Interface support depends on the client
- MCP does not provide sessions for external AI services
- Plugin visibility and identifiers remain unchanged