BoBman

Inspiration

BoBman was created to solve the "statelessness" problem in AI-assisted engineering. While AI coding assistants are powerful, they often lack a persistent "task brain" to track progress, dependencies, and complex engineering objectives across long-running sessions. The goal was to provide Cursor and Claude Code with a closed-loop, stateful orchestration layer that handles task decomposition and session persistence without requiring external APIs or incurring extra AI costs.

What it does

BoBman is a closed-loop engineering orchestrator that implements the Model Context Protocol (MCP). It functions as a stateful task manager for AI agents, allowing them to:

Decompose objectives: Break down free-form goals into a structured, numbered task graph (DAG). Maintain state: Track task progress, dependencies, and blockers in a local SQLite database. Analyze codebases: Use Tree-sitter to perform deep analysis (symbols, call graphs, and impact maps) across multiple languages including TypeScript, Python, Go, Java, and Rust. Assess risk: Generate risk scores for files and symbols based on churn, fan-in, and coverage gaps. Knowledge management: Provide a session-scoped knowledge base for facts, constraints, and decisions.

How I built it

Framework: Built as an MCP server using the official MCP SDK. Core Logic: A closed-loop orchestrator that manages a task graph, utilizing a stateful "get_next_task → work → report_complete" cycle. Database: Leveraged local SQLite for zero-cost, persistent storage of sessions, knowledge, and task graphs. Code Intelligence: Integrated Tree-sitter (with WASM grammars) for robust codebase parsing and analysis. Data & Metrics: Used simple-git for incremental indexing of code churn and Octokit for GitHub issue/PR context integration. Performance: Implemented Pino for logging and built support for both stdio and Streamable HTTP transports.

Challenges I ran into

Environment Compatibility: Managing native modules like better-sqlite3 across different Node versions (e.g., Cursor’s bundled Node vs. local system Node) required custom diagnostic tooling (npx bobman-mcp doctor) and specific documentation for Windows users. Concurrency: Handling SQLite locks when multiple processes or sessions access the database required managing WAL busy timeouts to ensure reliability. Tool Discovery: Ensuring that LLMs consistently utilize the tool suite required crafting imperative, behavior-driven tool descriptions that guide the agent to "check in" with BoBman at session start.

Accomplishments that I'm proud of

Closed-Loop Reliability: Successfully created a system where the agent transitions autonomously between DECOMPOSING, COMPLETE, and BLOCKED states without manual intervention. Feature Completeness: Mapped and implemented the PRD v1.0 and v1.1 feature set (F-01–F-11), ensuring professional-grade traceability. Intelligent Risk Scoring: Integrated actual coverage.json/lcov data into the risk scoring system, allowing the agent to make data-driven decisions about code changes.

What I learned

MCP Ecosystem: Deepened my understanding of the Model Context Protocol, particularly how to build stateful servers that feel like a native extension of the IDE/agent. Agent Behavior: Learned that effective agent-tool interaction is heavily dependent on the "imperative tone" of tool descriptions—the agent performs best when the toolset acts as a structured contract. Developer Experience (DX): Realized the importance of "Day 0" setup tools (like init and doctor commands) in ensuring successful adoption for complex node-based projects.

What's next for BoBman

PRD v2.0 Implementation: Expanding beyond the current scope to include Slack integration, a cloud dashboard, and automated PR bot capabilities. Extended Tooling: Continuing to refine the impact mapping and change hotspot analysis for even larger, enterprise-scale repositories. VS Code Adoption: Enhancing the read-only VS Code sidebar extension to further bridge the gap between MCP server data and the UI.

But This is not 100% done ,i am still working on it and making this more better . And i am more than happy to receive a contributions for others .

Thankyou..