Coredence

Coredence

Your mind's digital twin.

Coredence reads what you have already written, consolidates it into memories about how you actually work, and keeps every one of them tied to the traces it came from. The result lives in a single file on your own machine.

→ coredence.com · Try it in your browser — the real window on a made-up person's mind, nothing to install

What this repository is

Releases, documentation and the issue tracker. Not the source.

The engine that reads your work is not published. The reason, and what you get instead, is written out on coredence.com/license — but the short version is that the claims worth checking are checkable without it, and more conclusively:

lsof -i -a -p $(pgrep -x Coredence)      # prints nothing: the window opens no sockets
sqlite3 ~/.coredence/coredence.db '.tables'  # your whole mind, in a tool we did not write

Source code tells you what should happen. Those tell you what is happening.

One connection exists and it is worth stating plainly: while it is consolidating, the engine calls the model you chose — your Claude Code, your API key, or a server on your own machine. That is the only thing that leaves, it happens only during a cycle, and the last of those three options removes it entirely.

Status

Released. The current version is on the releases page, signed with a Developer ID and notarised by Apple.

Watch this repository (Watch → Custom → Releases) and GitHub will tell you when the next one lands. There is no mailing list, because there is no account and nothing to put you in one.

Getting it

One download from the releases page or from coredence.com — macOS 14 or later, Apple silicon. The engine ships inside the app — nothing to clone, no Python to install, no command to run.

On first launch it asks two things: where it should think, and what to call you. Reading your work needs a language model and the download does not include one, so you point it at the Claude Code you already have, at an API key of your own, or at any OpenAI-compatible server — including a model running on your own machine.

Windows works too — the engine, the installer and the window all ship. A mind is one file and moves between machines with you.

Already use an MCP client? Coredence is in the official MCP Registry as io.github.coredence/coredence. The server is read-only: it answers questions about a mind the app has already built on your machine.

Handing it to an assistant

The mind is more useful inside the tools you already work in than in a window of its own. Coredence ships an MCP server, so an assistant can ask what you think about a subject and what any single memory was drawn from, rather than being handed a file.

It runs on your machine — the client starts it and talks to it over a pipe, so there is no account, no key and nothing listening on a port. It is read-only, and that is mechanical rather than a promise: the connection opens the database read-only, so SQLite itself refuses a write. An assistant can quote your mind and disagree with it, but deciding who you are stays with you.

For Claude Desktop it ships as a desktop extension — coredence-<version>.mcpb on the releases page, five read-only tools:

The app connects the assistants it finds on your machine during setup. To add Coredence from an assistant instead, this repository is also a plugin and an extension. Install the app first, then open it once, so that the coredence command exists; the plugin only starts it with coredence mcp.

# Claude Code
/plugin install coredence --marketplace coredence/coredence

# Claude Code before 2.1.275: add the marketplace, then install
/plugin marketplace add coredence/coredence
/plugin install coredence@coredence

# Codex
codex plugin marketplace add coredence/coredence

# Gemini CLI
gemini extensions install https://github.com/coredence/coredence

VS Code and GitHub Copilot read the same plugin (plugin.json, mcp.json). What the plugin runs is the read-only server above and nothing else; it carries no code of its own.

From a terminal, coredence connect <client> writes the configuration for you — Claude Code, Cursor, VS Code, Windsurf, JetBrains, Codex and Gemini CLI are known by name.

Assistants do not always call an MCP server on their own, so the mind can also be handed over directly:

• As an instructions file — Codex (~/.codex/AGENTS.md), Gemini CLI (~/.gemini/GEMINI.md) and VS Code with Copilot (~/.copilot/copilot-instructions.md) read one every session. Coredence writes its own marked section and refreshes it every cycle, leaving the rest of the file alone. Settings › Assistants, or coredence instructions <name> --on.
• To chat assistants — Claude, ChatGPT and Gemini chats take it through their own memory import or custom instructions. Settings › Assistants › Copy and open copies your strongest memories at the length each one accepts and opens the right page; you paste. Nothing is sent from the app.

Documentation

• Getting started — what it reads, what the five layers mean, what to do with a memory you disagree with
• Privacy — what is collected (nothing) and how to verify that rather than believe it
• Security — signing, notarisation, and how to check a build before you open it
• Roadmap — what is next, and what has been decided against

Issues and questions

Both belong here. Nothing is triaged in private, and there is no support address that goes somewhere different from where the work is.

For anything that should not be public, open a private security advisory instead.

Made by Akaya Labs.