CORE is a self-hosted personal assistant that runs continuously rather than being opened as a chat window. It indexes email, meetings, GitHub, Linear, Slack, and assistant conversations into a temporal knowledge graph, then starts each unit of work with that context already loaded.
Work is organised as tasks, each carrying a specification, a generated plan, live state, and its own chat thread; tasks can be one-shot or recurring, and each can spawn a coding session, drive a browser, or run terminal commands. A daily scratchpad page is the usual entry point — a checkbox line is picked up within about three minutes, context is gathered, and a plan is drafted for approval. Other entry points include a push-to-talk voice shortcut on macOS and messaging through WhatsApp, Slack, or Telegram, all sharing the same memory.
Coding work is delegated to Claude Code or Codex sessions started with the full task context, running on the local machine or in Docker or Railway, continuing while the laptop is closed, and opening a pull request when finished. Rather than a separate API key, CORE can be pointed at an existing Claude or Codex subscription through a proxy, which also covers memory, planning, and task execution. A built-in browser, isolated from the user's main browser, limits which accounts the assistant can reach.
Integrations arrive through MCP: one-click connections to more than 50 applications and roughly 1,000 actions, with webhook triggers turning inbound events into new work. Installation is a Docker-based stack started by a setup wizard on localhost:3033, or a one-click Railway deployment, with a gateway connected separately so the assistant can run coding agents, drive the browser, and reach local folders. Requirements are Docker 20.10+, Docker Compose 2.20+, and 4 vCPU with 8 GB of RAM. On the LoCoMo memory benchmark the project reports 88.24% average accuracy across single-hop, multi-hop, open-domain, and temporal reasoning.
Features
- Temporal knowledge graph: email, meetings, GitHub, Linear, Slack, and chat history indexed as memory with fact classification and intent-driven retrieval
- Tasks: spec, plan, live state, and a dedicated thread per unit of work, one-shot or recurring
- Scratchpad: a daily page where a plain checkbox line becomes a picked-up task within minutes
- Voice: a keyboard shortcut on macOS starts a background task without switching windows
- Messaging gateway: reachable from WhatsApp, Slack, Telegram, email, web, and API with shared memory
- Coding sessions: Claude Code or Codex runs on your machine, in Docker, or on Railway, ending in a pull request
- Subscription proxy: the stack runs on an existing Claude or Codex subscription instead of per-token API billing
- Isolated browser: a built-in browser the assistant drives, logged in only to the accounts you choose
- Connectors: MCP integrations to 50+ apps and around 1,000 actions, with webhook triggers
- Skills: reusable instructions for repeatable workflows
- Human-in-the-loop: plans are approved before they run
- Self-hosting: Docker Compose stack or one-click Railway deployment, with data kept in your own infrastructure
