Cortex - RAG MCP for a knowledge base

English | Français

Cortex is an MCP (Model Context Protocol) server that exposes semantic search over a local knowledge base. It lets Claude, Codex and Gemini find the right passage in your documents without wasting their context window. Search is semantic (by meaning, not keyword), in French and in English. Cortex processes and indexes the knowledge base locally without sending its content; the MCP client may still pass requested chunks to its model under its own policy. The optional Confluence writer only downloads explicitly allowlisted spaces; the generated Markdown, vector index, and lexical index remain local.

Starting with 2026.0906.01, Companion offers a guided home screen, operation history and Confluence setup from one page or space link. Connection and measured scope confirmation stay in the flow; successful collection is followed by indexing. Use the combined installer to keep Cortex and Companion compatible.

With paired version 2026.0909.02, Confluence updates preserve the served generation when subtree enumeration fails and apply changed target/classification settings on retry. Search uses multilingual vector retrieval by default; CLI and MCP callers can explicitly select hybrid or reranked search. The first successful collection after upgrading regenerates older publication revisions, and the following sync removes obsolete index entries.

Guided document updates (2026.0909.00)

Companion brings local documents and Confluence together in My sources, with one Update my documents action. Search previews are resizable, large page trees render visible rows, and cancelling review preparation keeps the draft. Readiness now follows the saved document folder and configuration, including after restart. Run one document update after upgrading to establish this evidence for older indexes. See the release notes.

Measured scope without enumerating the space (2026.0907.00)

Adding a Confluence source measures its page, subtree and whole-space scopes with one indexed count each instead of reading the space page by page. On a 5916-page space that measurement took 3 min 19 s, past the timeout of every graphical caller, so Companion showed nothing at all. It now answers in about a second, and the emitted contract is unchanged.

Two whole-space numbers shift by one as a result: the resolved page is no longer folded into the space total, and a space with no visible page reports zero rather than one. A deployment whose search endpoint returns no total cannot measure a scope and reports that as a permanent failure instead of a retryable one. See the contracts.

Paired Companion 2026.0907.00 restores the contrast of the scope window, whose options were drawn in the system text colour, and stops reporting a timeout as a connection failure or advising a delay increase that silently reverts.

Source management in Companion (2026.0906.02)

The workflow includes search, remote page trees, impact review, evidence-based readiness, save-and-update, last-removal undo and recovery actions. See the contracts.

Paired version 2026.0906.02 provides visible My sources cards, a prefilled selection editor, confirmed page/space removal and browser links to originals. Removal only changes Cortex tracking; search reflects it after successful collection and indexing.

Removing the final source explicitly writes spaces = [] in a schema v2 or v3 configuration. This intentional empty allowlist permits an empty publication with tombstones for prior documents. An absent spaces key remains incomplete configuration and collection is refused. Connection and credential validation and publication safeguards still apply. Both updated components are required; this capability is absent from the installed 2026.0906.01 release.

Installation

Windows, no Python (recommended)

The simplest path: one installer for Cortex, Cortex Companion, the windowless Confluence converter, and the offline models. No separate Python or .NET runtime is required.

1. Download Cortex-Setup.exe and SHA256SUMS from the latest release.
2. Before running the unsigned installer, calculate its digest with Get-FileHash .\Cortex-Setup.exe -Algorithm SHA256 in PowerShell and verify that it exactly matches the Cortex-Setup.exe line in SHA256SUMS.
3. Double-click only after that check. If SmartScreen still warns, select More info, then Run anyway.
4. Choose the folder that holds your documents, keep Index everything in this folder, and finish. Cortex Companion opens when installation completes. Indexing only starts when you request synchronization in Companion.
5. In Companion, open Settings and verify the knowledge-base folder. The Cortex executable installed with Companion is detected automatically.
6. Drop your documents in that folder, open Local database, then select Synchronize local documents.
7. Restart your AI application: Cortex shows up there as an MCP server.

Companion then lets you synchronize, schedule, diagnose, and configure Cortex without a terminal. Details, silent mode and reinstall: Windows install.

Standalone archives (Windows x64, macOS Apple Silicon, Linux x64)

Every release also ships one ZIP archive per platform. It contains the single cortex or cortex.exe binary (MCP server + CLI, no Python) and the licenses for every embedded dependency. See Standalone distribution.

From PyPI (Python, advanced)

py -m pip install --upgrade cortex-local-rag
cortex setup

This path installs the CLI and MCP server, but not Cortex Companion. The model is downloaded on first use if its cache is empty.

From source (Python, advanced)

:: From the folder where you cloned Cortex
install.bat

install.bat initializes the configuration, installs the dependencies, offers to register Cortex in the detected MCP clients, and validates the installation. Details: Setup.

How it works

Documents folder (.md, .pdf)       Optional Confluence writer (REST)
      |                                      |
      |                              current Markdown generation
      +------------------+-------------------+
                         |
                         v
  cortex sync           <- Split, hash, vectorize, update FTS5
                         |
                         v
  %LOCALAPPDATA%\Cortex\  <- ChromaDB + lexical.db
      |
      v
  cortex serve          <- MCP server (FastMCP)
      |
      v
  MCP clients           <- Claude / Codex / Gemini / Antigravity / LM Studio / Cursor / Windsurf / VS Code

The embedding model is the multilingual ONNX paraphrase-multilingual-MiniLM-L12-v2. The Windows installer bundles it; a source installation or standalone binary downloads it when the local cache is empty.

Two indexing modes

• Whole folder (default): anything you place in the chosen folder, at the root or in any subfolder, becomes searchable. Nothing to configure.
• Sections (advanced): limits indexing to named subfolders you can search separately (defaults knowledge, projects, notes).

These modes govern the user-selected document folder. Generated ingestion documents are indexed separately from the current published generation with source_kind=doc and section sources.

Details: Configuration.

The cortex command

The installed package exposes a single command:

cortex --help describes every subcommand and cortex <command> --help describes its options. A prompt-free installation scripts as:

cortex setup --yes --kb-path "D:\Documents\Knowledge"

Exposed MCP tools

Documentation