BRIEF MCP

Landing Page
MCP
Token Savings
Knowledge Base
Configuration

The Targeted Problem

Teams (and especially dev agencies juggling multiple client projects) get wrecked by context. Specs live in giant markdown files across random folders and tools, and every AI agent ends up with a different slice of “truth.” So people do the worst workflow: paste 1,000+ lines of PRDs/specs into prompts, burn credits, and still get inconsistent answers. The outcome is duplicated work, decisions getting re-litigated, and a lot of “wait… didn’t we already decide this?”

The Proposed Solution

BRIEFMCP is an MCP server that gives every AI agent the same, up-to-date project context without paste dumping massive markdown files. Inspired by John Lam’s spec first development approach, we structure knowledge into a Constitution (principles, constraints, conventions) and Feature Specifications (requirements, acceptance criteria, edge cases). BRIEFMCP indexes these into a knowledge base and uses RAG to route a question to the right sections, retrieve only what matters, and compress it into a small, cited Brief Packet under a token budget. Agents can query the spec directly and get the essentials with receipts, so teams stay aligned and stop nuking their context window.

The Technical Stack

We built BRIEFMCP with Next.js on the frontend and Supabase on the backend using Postgres with pgvector and Edge Functions. We generate embeddings with OpenAI and use a Featherless hosted LLM for routing, packetization, and answer generation. Everything is delivered via MCP so compatible assistants can fetch context on demand. The core pipeline is route, retrieve, compress, then answer from the packet.

Limits & Avenues for Development

Right now the MVP focuses on markdown ingestion and the MCP workflow, with manual uploads and limited integrations. Next we want to add connectors for Notion, Confluence, Slack, and Figma, introduce a first class Decision layer plus “what changed since yesterday” context deltas, improve retrieval with hybrid keyword plus vector search and better deduping, and harden the system for enterprise with permissions, audit logs, and prompt injection safety.