alexlabs-ai/brain-concierge

Brain Concierge is a retrieval layer that sits between an AI agent and a knowledge base. The README opens with the problem it is built around: an agent searching for a single phrase like 'refund policy' will never think to also search for 'VIP exceptions' or 'January outage compensation', because it does not know those concepts are in the corpus. Standard search puts the burden on the asker to phrase the right query. The pattern is task-first rather than search-first. Instead of sending a keyword, the agent describes what it is about to do, optionally including a role like 'account executive' or 'backend engineer'. Brain Concierge then expands that description into several semantic queries, runs them in parallel against a GBrain vector index, deduplicates and re-ranks the results, and asks an LLM to synthesize a briefing focused on the task. Every response comes back in three parts: the briefing itself, the source pages that fed it with slugs for traceability, and a small KB index showing corpus stats and the exact queries that were run. The README gives concrete examples. A sales agent preparing for a renewal call describes the customer situation and goal; an engineering agent about to migrate from session auth to JWT describes the change and the risk it is worried about; a market analyst writes a brief about entering a healthcare compliance market. There are optional parameters for depth (standard or deep) and for restricting results to specific slug prefixes when the knowledge base is partitioned by topic. Installation needs Node.js 18 or newer, a running GBrain server, an Anthropic API key for the synthesis step, and an MCP-compatible client such as OpenClaw, Claude Desktop, or Cursor. You clone the repo, install, fill in .env, and start the MCP server on port 7351. By default it uses claude-sonnet-4-6 for synthesis, and the model can be swapped via the SYNTHESIS_MODEL environment variable. The repository ships Promptfoo eval configs, with a candid note that benchmarking task-first retrieval against query-first retrieval has a structural limit because anyone writing test queries already knows the corpus. The project is MIT licensed, version 1.1.0, and built on top of GBrain by Garry Tan.