lotus-wisdom-mcp is an MCP (Model Context Protocol) server implementation that provides problem-solving tools based on the wisdom framework of the Lotus Sutra, combining analytical thinking with intuitive wisdom. The project evolved from the Lotus OS prompt, and its MCP server form makes the framework easier for Claude and other AI assistants to use.
[Main Capabilities] It adopts a multi-faceted problem-solving approach inspired by the Lotus Sutra, including a step-by-step thinking process and different thinking techniques, and adds meditative pauses to let insights emerge naturally. It supports interactive visualization through MCP ext-apps (adapting to the host's light and dark themes, keyboard accessible). It provides the MCP prompts contemplate and deep-inquiry for one-step guided contemplative sessions. In addition to text responses, it also provides structured tool output (structuredContent + outputSchema). It simultaneously tracks the tag journey and wisdom domain movements, ultimately integrating insights into a clear response.
[Wisdom Domains and Tags] The server uses wisdom domains to organize thinking. All valid values for the tag input parameter are as follows: Entrance (🚪): begin. Start the journey here; the complete framework is received before contemplation begins. Skillful Means (🔆): upaya, expedient, direct, gradual, sudden. Different methods leading to truth, sometimes pointing directly, sometimes unfolding gradually. Non-Dual Cognition (☯️): recognize, transform, integrate, transcend, embody. Awakening to the various aspects of what already exists; recognition is transformation. Metacognition (🧠): examine, reflect, verify, refine, complete. The mind observes its own understanding gradually unfolding. Process Flow (🌊): open, engage, express. A natural arc that can accommodate any of the above methods. Meditation (🧘): meditate. Pause, and let insight emerge from silence.
[Process] The user submits a problem to be solved; the model starts with tag='begin' to receive the complete framework; then continues using contemplative tags (open, examine, integrate, etc.); each thought builds on the previous one and may revise understanding; the tool tracks the tag journey and wisdom domain movements; meditative pauses can be added to maintain clarity; when status='WISDOM_READY' is returned, the tool's work is complete; the model thenThen express the final wisdom naturally in your own voice.
[Available Tools] lotuswisdom_contemplate: Solve problems based on the wisdom framework of the Lotus Sutra. Begin the journey with tag='begin', returning the complete framework (philosophy, domains, guidance) to lay the foundation for contemplation, then continue using other tags. Parameters: tag (string, required, the currently processed technique), content (non-empty string, required, the content of the current processing step, including begin), stepNumber (integer, optional, default 1, the current sequence number), totalSteps (integer, optional, default 5, the estimated total number of steps), nextStepNeeded (boolean, optional, default true, whether the next step is needed), isMeditation (boolean, optional, whether it is a meditation pause), meditationDuration (integer, optional, meditation duration in seconds, 1-10), previousJourney (string, optional, the journey string from the previous response, e.g. "begin → open → examine", used for stateless clients to continue the journey). Returns a JSON text block and structuredContent validated by outputSchema. Response statuses include: processing status with current step information, wisdom domain, and journey tracking; FRAMEWORK_RECEIVED status for the begin step; MEDITATION_COMPLETE status for the meditation step; WISDOM_READY status when the contemplation process is complete. This tool declares readOnlyHint, idempotentHint, destructiveHint: false, openWorldHint: false. These are host hints and do not guarantee zero side effects: local stdio calls update the in-memory journey, and hosted workers log usage analytics. The tool does not modify user files or external business data. lotuswisdom_summary: Get a summary of the current contemplation journey. Parameter previousJourney (string, optional, the journey string from the previous response, used to reconstruct the summary on stateless clients). Returns the domain journey showing movement between wisdom domains, as well as a summary of all steps with their tags, domains, and brief content.
[MCP Prompts] contemplate (parameter question): Start asingle-question contemplation, instructing the model to begin with tag='begin', iterate, and speak the wisdom only when status='WISDOM_READY'. deep-inquiry (parameter topic): Begin a longer inquiry, consciously moving across wisdom domains (process → metacognition → non-duality → meditation).
[Applicable Scenarios] Decompose complex problems that require multi-faceted understanding; problems that benefit from both direct and gradual approaches; problems that require integrating surface contradictions; situations that require analytical and intuitive understanding; tasks that benefit from meditation pauses to let insights emerge; problems that inherently contain inner wisdom.
[Installation and Usage] Can be installed with one click via Smithery, or configured manually. Requires Node.js 18+, and the server runs locally via npx. CLI installation (one-line command). claude_desktop_config.json, .vscode/mcp.json (VS Code uses "servers" as the top-level key), ~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, the Cline panel, Settings > MCP Servers, etc. all have corresponding configuration methods. Codex CLI (TOML configuration) and Gemini CLI (JSON configuration) can edit the corresponding configuration files. On Windows, npx requires a shell wrapper, and "command": "npx" needs to be replaced. ChatGPT only supports HTTPS remote MCP servers; you can use Smithery or connect directly to the hosted instance described below via Settings > Connectors. Remote (hosted): the public instance is located at https://lotus-wisdom-mcp.linxule.workers.dev/mcp , no API key required. Clients that support Streamable HTTP can connect directly; clients that only support stdio can use mcp-remote. For self-hosting, see worker/README.md. Building from source: use Bun 1.4.2 and Node.js 24 to match CI. The build-install app locks dependencies and rebuilds the dist/bundle.js and dist/journey.html artifacts. App type checking and build regression: cd app && bunx tsc --noEmit && bun run test. Dependabot uses the bun ecosystem to handle the root, app, and worker packages. Merging a PR only updates the source: version tags are published to npm and the MCP Registry, and the Cloudflare Worker must be deployed separately. Debug mode: when LOTUS_DEBUG=true is enabled, the local stdio server can output per-step trace lines to the console (stderr).
[Interactive Visualization (ext-apps)] In MCP clients that support ext-apps (Claude Desktop, Cursor, ChatGPT), the tool renders an interactive "Living Trace" visualization inline in the chat: journey trail (SVG circles colored by wisdom domain), domain colors (process gold, skillful means amber, non-duality green, metacognition blue, meditation cyan), meditation breathing (hollow circle with gentle inhale/exhale animation), completion (the journey resolves into a gradient path showing the complete domain arc), click to explore (pin any step to read its contemplation text), long journey folding (show the most recent 8 steps, with earlier steps grouped as "+N"). Clients that do not support ext-apps are unaffected and still receive the same JSON tool response.
[Token Optimization Design] MCP tool descriptions always remain in the AI context window when the server is connected. To minimize overhead while retaining the full teaching content: constant context (about 150 tokens), the lotuswisdom tool description stays concise; on-demand learning (about 1,200 tokens), calling tag='begin' delivers the complete framework, including philosophy and domain spirit, parameter descriptions (tag, content, stepNumber, etc.), response format details (wisdomDomain, journey, domainJourney), meditation handling (MEDITATION_COMPLETE status), and guidance on when to use it. Learn first, then practice: the begin tag ensures the model gains complete understanding before contemplating. This approach reduces constant context overhead by about 85% when the tool is idle.
[Version and License] Current version 0.8.1, adding updated dependencies and GitHub Action. This MCP server uses the MIT license. Contributions are welcome.
