MCP and plugins

MCP servers

Settings → MCP Servers connects Model Context Protocol servers, such as Blender’s or your editor’s, so agents can use their tools.

• Local servers start as a command with arguments and environment variables, the same way Codex or Claude Desktop starts them. Remote servers are a URL.
• Import reads the servers you already configured in Codex (~/.codex/config.toml), Claude Desktop or Cursor, or a pasted snippet.
• Approval is set per server: ask on every call (the default), approve read-only tools automatically, or trust the server.
• In workflows, the MCP tool step calls one tool directly, without a model.

Tool results that contain images reach only models that can see images.

Plugins

Settings → Plugins lists Notion, GitHub, Slack, Google Drive, Linear, Jira, Outlook Email, Outlook Calendar, Microsoft Teams and Dropbox.

1. Install the plugin.
2. Connect an account in your browser.
3. Choose Read only or Read & write, and enable it.

Installing alone grants no access. Every write asks for approval, even with full access. Account tokens stay on your device, encrypted, and never reach prompts or tool output. Settings → Connected accounts lists the accounts; disconnecting deletes the local credentials.

Local Cognitive as an MCP server

Other agents can call Local Cognitive’s models through MCP over stdio. From a source checkout:

npm run build
npm run mcp:stdio

It offers local_ai_chat, local_ai_code, local_ai_hypothesis, local_ai_runtime_status, local_ai_list_models, local_ai_get_session_settings and local_ai_update_session_settings.