IT-MCP

<img src="https://r2cdn.perplexity.ai/pplx-full-logo-primary-dark%402x.png" style="height:64px;margin-right:32px"/> # Local File-Based Structured Thinking MCP: Advanced Planning Architecture This guide details how to establish a completely local, file-based Model Context Protocol (MCP) system supporting advanced structured planning. It leverages a SQLite database with jsonb-emulating capabilities and FTS5 (for near-GIN speeds) for all persistent state, rooted at the workspace. The system bootstraps from user and workspace `.md` knowledge resources, ensuring planning is both systematic and fully local and auditable.[^1][^2][^3][^4][^5][^6] *** ## Architecture Overview The local MCP stack centers on three core principles: - **Local-only state**: All planning data (thoughts, plan objects, reasoning state, file metadata) is persisted in a local SQLite file (e.g., `mcp_plan.db`) at the workspace root. - **Systematic bootstrap**: On initialization, the agent scans the workspace for all `.md` resources (such as `agents.md`, `CLAUDE.md`, architecture docs), auto-parses them, and encodes summaries/metadata for structured query. - **Advanced planning focus**: The system is designed for medium and long-term project planning; prompt-level short-term thinking should be filtered out. Each stored plan object is structured as a JSON blob, referencing the source `.md` files and maintaining full auditability. *** ## Local Server Deployment ### SQLite MCP Server Deploy a SQLite MCP server using the Node.js or Python implementations:[^2][^4] - Add to your project via a configuration file (e.g., `.vscode/mcp.json`, `claude_desktop_config.json`): ```json { "mcpServers": { "sqlite": { "command": "npx", "args": ["-y", "mcp-server-sqlite-npx", "${workspaceFolder}/mcp_plan.db"], "type": "stdio" } } } ``` - This ensures all queries and persisted state are local, with zero cloud dependency. *** ## Workspace Bootstrapping Workflow 1. **File Discovery**: - Agent uses filesystem MCP or native server tools to enumerate `*.md` files in the workspace root.[^7][^5][^8] 2. **Markdown Parsing**: - Each `.md` file is parsed by section and key metadata is extracted (e.g., filename, title, update time, tags). - Parsed content and outlines are stored as JSON documents in SQLite database for rapid semantic access. 3. **Systematic Review**: - Agent creates structured summaries for each imported `.md` file, tagging resources for traceability and reasoning.[^6] 4. **Structured Agent Initialization**: - Agent uses both the parsed `.md` resources and user-provided input to synthesize new planning objects.[^9][^5][^8] - These objects explicitly encode: - Planning horizon (medium/long-term) - Stage/context (e.g., upgrade roadmap, change window) - Linked evidence from workspace files - Reasoning score for metacognitive feedback 5. **Plan Storage**: - Each plan/reasoning object is stored as a JSON document in the local database. - Audit trail includes links to all referenced `.md` resources and encoded user/system input. *** ## JSON Schema for Thought Object ```json { "thought_id": "uuid", "type": "planning", "horizon": "medium-term|long-term", "topic": "Infrastructure upgrade roadmap", "inputs": { "user": "Alex Campkin", "resources": ["agents.md", "CLAUDE.md"] }, "imported_md": { "agents.md": { "title": "Agent Decision Logic", "sections": ["Goals", "Stages", "Criteria"], "tags": ["planning","ops","automation"], "latest_update": "2025-10-29T15:00:00Z" } }, "context": { "stage": "Advanced Planning", "domain": "devops", "window": "2025 Q4 - 2026 Q3" }, "plan_details": { "objectives": ["Upgrade", "Risk mitigation", "Automate"], "subgoals": ["Test resilience", "Document runbooks"], "review": "Synthesis of .md docs and user priorities" }, "links": ["./agents.md#logic", "./network.md#roadmap"], "score": 0.92 } ``` *** ## Advanced Planning Stages Supported - **Objective Cataloging**: Extract projects and goals from all workspace docs. - **Constraint Analysis**: Parse operational/compliance constraints from `infrastructure.md`, `security.md`, etc. - **Milestone Planning**: Encode and track major deliverables/milestones. - **Branching/Revision**: Use MCP branching tools to model alternative strategies and what-if scenarios.[^5][^9] - **Audit/Traceability**: Capture all logic, file provenance, and decision reviews for downstream or compliance checks. *** ## Database Management Notes - SQLite’s `json` and `fts5` modules allow efficient index/query, giving near-GIN performance for local setups.[^4][^2] - Ensure that `mcp_plan.db` is always at workspace root for portability and rapid sharing/versioning. - Back up local plans manually if needed—no state is sent to the cloud by default.[^5] - Protect database location via strict OS/FS permissions and MCP server config: restrict tool access to just the workspace. *** ## Bootstrap and Best Practices - On each agent run, parse and catalog all `.md` files for structured review.[^7][^6][^5] - Only permit workspace-folder access when configuring your MCP server—never expose system directories by default. - Each advanced plan must reference `.md` origin data in its JSON structure for proper lineage and audit. - Score and tag all plans for review and future upgrades. - Use native MCP query and search tools to dynamically update, review, and branch plans as new files or requirements are added. *** **Version**: 1.0 **Last Updated**: 2025-10-29 **Maintainer**: Workspace Owner <span style="display:none">[^10][^11][^12][^13][^14][^15][^16][^17][^18][^19][^20]</span> <div align="center">⁂</div> [^1]: https://skywork.ai/skypage/en/unlocking-local-data-sqlite-explorer/1978663309412704256 [^2]: https://github.com/makeralchemy/claude-desktop-mcp-sqlite [^3]: https://github.com/cnosuke/mcp-sqlite [^4]: https://playbooks.com/mcp/johnnyoshika-sqlite-npx [^5]: https://skywork.ai/skypage/en/unlocking-structured-ai-reasoning/1977642632387035136 [^6]: https://taoofmac.com/space/blog/2025/10/04/1111 [^7]: https://skywork.ai/skypage/en/mark3labs%2Fmcp-filesystem-server: A Deep Dive for AI Engineers/1971095480809418752 [^8]: https://playbooks.com/mcp/safurrier-filesystem [^9]: https://github.com/Promptly-Technologies-LLC/mcp-structured-thinking [^10]: https://skywork.ai/skypage/en/sqlite-mcp-server-guide-ai-engineers/1977629649939787776 [^11]: https://playbooks.com/mcp/isaac-gounton-sqlite [^12]: https://skywork.ai/skypage/en/cheny-alf-secure-filesystem-mcp-server/1977922670831783936 [^13]: https://mcpservers.org/servers/panasenco/mcp-sqlite [^14]: https://skywork.ai/skypage/en/phillips-think-tool-ai-engineers-guide/1977622899914641408 [^15]: https://n8n.io/workflows/3632-build-your-own-sqlite-mcp-server/ [^16]: https://modelcontextprotocol.io/examples [^17]: https://github.com/modelcontextprotocol/servers [^18]: https://dev.to/copilotkit/30-mcp-ideas-with-complete-source-code-d8e [^19]: https://github.com/apappascs/mcp-servers-hub [^20]: https://www.reddit.com/r/ClaudeAI/comments/1jf4hnt/setting_up_mcp_servers_in_claude_code_a_tech/