← all projects
OffboundRAG & MCP

An AI that answers from your own knowledge.

For Offbound, I deployed a RAG that indexes all of the company's data and that of its Skool community. An MCP server opens it up to Claude: each Offbound member will be able to connect Claude to it and directly query the company's knowledge.

// the problem

Rich knowledge, split between the company and the community.

01 · scattered content

The company's documents on one side, the Skool community's modules and lessons on the other, each with its own format.

02 · searching by hand

Finding the right lesson requires already knowing where it is.

03 · generic AI

A general-purpose assistant answers with what it knows from the web, without Offbound's own method.

// what I've built

All of Offbound's knowledge, queryable from Claude.

Three separate building blocks: the base keeps the raw data, the RAG only knows what is pushed to it, the AI only talks to the RAG.

stack →SkoolSupabaseOpenRouterTypeScriptMCPClaude
  1. 01

    A notes base

    The company's data and the Skool community's content are stored as Markdown notes, in folders. New notes also arrive through an API entry point, for n8n, Zapier or Make.

  2. 02

    Chosen indexing

    Storing a note doesn't index it: it enters the RAG through an explicit action, note by note or in batches. A note modified after its indexing is flagged as outdated.

  3. 03

    A hybrid RAG

    The search combines meaning and exact words, filters by module or by folder, and returns one result per document, represented by its best excerpt.

  4. 04

    An MCP server for each member

    Each member connects Claude to the MCP server with their own revocable token. Claude searches, reads and lists the indexed documents, without modifying anything.

$ search_documents "structuring an offer" · module: offer
→ hybrid search · meaning + exact words
→ one result per document · best excerpt
$ read_document · full text
✓ answer written · source cited

// the workflow

Companyinternal documents
Skoolmodules and lessons
Indexingexplicit action
RAGmeaning + exact words
a_
Claudeeach member via MCP
Answersource cited
✓ sourced answercompany + Skoolone token per member
the principle of the workflow · simplified diagram

// result

Sourced answers

Offbound's knowledge fits in a single index. Claude answers from the content of the company and the community, and cites the document each answer comes from.

// projects

The other projects.

All projects →
Simplement · AI agent

Proposal writing agent

An application where sales reps prepare their proposals with an AI agent.

Korya · AI agent

Email sorting agent

Every morning, sorts emails by importance and prepares the action to take for each important email.

Korya · Web agency

Internal cockpit driven from Claude

Automated onboarding, production and reporting: 2 to 3 days gained per week.

// contact

An AI that answers from your knowledge? Let's talk about it.