What ConnectAI is

Introduction

ConnectAI is self-hostable company-brain infrastructure. You connect the tools a business already uses, ConnectAI derives a durable company brain from them, and it serves that brain to any agent over a read-only MCP server. You run all of it yourself.

flowchart LR
  accTitle: ConnectAI architecture
  accDescr: Connected sources flow through connectors and ingest into the company brain (Postgres plus pgvector), which a read-only MCP server serves to your agents.
  S["Connected sources<br/>email, calendar, chat, CRM, docs"] --> I["Connectors + ingest"]
  I --> B["Company brain<br/>Postgres + pgvector"]
  B --> M["Read-only MCP server"]
  M --> A["Your agents<br/>Claude, Cursor, CLI, apps"]

The mental model in one minute

The product is a substrate, not an app. It has two parts, and both are self-hostable:

Part	What it is
The database	The governed-raw company brain: Postgres plus pgvector (1024-dim embeddings, HNSW indexes, row-level security, an append-only action ledger). Connectors ingest the tools a business runs on; a brain pipeline synthesizes, embeds, and serves retrieval over the full substrate.
The MCP server	A remote, OAuth-authenticated, read-only Model Context Protocol endpoint. You point your own agent (Claude, Cursor, a CLI, a custom app) at your brain, scoped by row-level security to your own workspace.

You build your agents and apps on top. ConnectAI owns the connection sessions, the derived brain, the read-only MCP surface over it, and the action ledger. The source systems you connect stay the systems of record: ConnectAI keeps a governed, durable copy of the chunked raw to give you maximum recall and linking, not a normalized mirror of every vendor record.

What "governed-raw" means

ConnectAI durably stores chunked raw content from your connected sources, per workspace, encrypted at rest in a key-separated zone. The trust claim is precise:

Encrypted, workspace-isolated, never trained on, retention you control, and one-tap purgeable. It is not "we never hold your raw."

Holding the raw is what lets retrieval run over the full substrate instead of a lossy distillation, and it lets an answer stay grounded even after the upstream source is deleted.

The five-minute pitch

A business already describes itself every day, in email, calendar events, chat threads, meeting notes, CRM updates, issues, and invoices. ConnectAI reads those sources through connectors, builds one queryable brain from them, and exposes that brain to your agents over a standard protocol. You self-host the database and the MCP server, so the company's most sensitive intelligence never leaves your infrastructure.

A note on scope

ConnectAI also ships a reference proof-of-concept: a continuous per-business loop that senses (ingests), derives (synthesizes the brain), reasons, proposes an insight plus a drafted action, acts on one-tap owner approval, and learns from the owner's edits. Treat that loop as a reference built on the substrate, not the headline. The durable product you self-host is the database plus the read-only MCP surface over it.

ConnectAI is not a task-runner, a co-pilot, or a workflow builder. It is the queryable brain your own agents read from.

Where to go next

Quickstart runs the one-command installer and gets you to a live, queryable brain in a few minutes.
Self-hosting (advanced) covers the manual docker-compose path, environment, vault, and model right-sizing.
Connect your agent (MCP) points Claude, Cursor, or a CLI at your brain over the read-only MCP endpoint.