Today you use specialized agents like Claude Code and Cursor. You gather the context, kick off the session, babysit the output. Soon you will just describe the outcome, and Butler runs Claude Code, pulls from your apps, coordinates the agents. You stop being the orchestrator. Butler becomes the harness.
core-demo-github.mp4
Why we're building this
Every AI agent you use today is smart. And every single one forgets you the moment the conversation ends.
Chat is the wrong interface for a personal assistant.
Your EA doesn't wait for you to open a chat window and explain what you need. They already know. They're already on it. Chat forces you to context-switch, explain yourself, and stay in the loop on things that shouldn't need your attention.
We think the right interface is a scratchpad, a shared page, like a note you and a colleague both have open at the same time. You write what's on your mind: tasks, thoughts, half-formed ideas. Butler reads it alongside you, picks up what's meant for it, works in the background, and updates you when something needs your attention or a job is done. No prompting required, no workflow to configure. When you need to go deep on something specific, tasks and direct chat are there, but the default should be the scratchpad.
Your AI doesn't actually know you.
Every agent starts fresh. No preferences, no past decisions, no team context, no patterns. So it can't act proactively, because proactivity requires context that accumulates over time. Without persistent memory, agents are reactive tools waiting for your next prompt.
You are the bottleneck.
Right now you are the glue. You gather the GitHub issue, read the Slack thread, check the error logs, paste it all into Claude Code, wait for output, and review it. That's not delegating, that's operating as an agent yourself. Every workflow starts with you. Every session depends on you explaining things from scratch. CORE is built to break that loop.
Butler in action
You closed your laptop. Your meeting already produced results.
Butler read the transcript, extracted the follow-ups, created the tasks, and drafted the emails. You open the scratchpad and review. Done in four minutes.
You wrote one line. You came back to a PR.
[ ] fix the checkout bug from issue #312 — Butler loaded the context, spun up a Claude Code session, and handled it. You never left the terminal.
You opened your inbox to review, not to triage.
Butler flagged what needs you, drafted replies for the rest, and turned action items into tasks. Your inbox is a decision queue now.
You slept. The Sentry alert got handled.
CORE investigated, created the issue, and assigned the right engineer. You woke up to: "Handled. Issue #847, assigned to Harshith."
What CORE is not
| Not a RAG wrapper. | Memory isn't "embed chunks and search." It's a temporal knowledge graph where facts are classified, connected, and updated over time. It knows when you decided something and why. |
| Not a workflow builder. | No drag-and-drop. You write what needs doing. Butler figures out the workflow. |
Quickstart
Open source, self-hosted. Your data never leaves your infra.
git clone https://github.com/RedPlanetHQ/core.git cd core/hosting/docker cp .env.example .env # Add your model key (see below) docker compose up -d
Open http://localhost:3000 → connect your first app → hand off your first task.
Pick your model (edit .env):
OPENAI_API_KEY=sk-... # OpenAI ANTHROPIC_API_KEY=sk-ant-... # Anthropic Claude OLLAMA_URL=http://localhost:11434 # Ollama — fully local, no cloud OPENAI_BASE_URL=https://your-proxy.com # Any OpenAI-compatible endpoint
Requirements: Docker 20.10+, Docker Compose 2.20+, 4 vCPU / 8GB RAM
☁️ Prefer cloud? Try CORE cloud free, 3,000 credits included. Get started →
Benchmark
CORE achieves 88.24% average accuracy on the LoCoMo benchmark — single-hop, multi-hop, open-domain, and temporal reasoning.
Docs
Want to understand how CORE works under the hood?
- Memory — Temporal knowledge graph, fact classification, intent-driven retrieval
- Toolkit — 1000+ actions across 50+ apps via MCP
- CORE Agent — The orchestrator: triggers, memory, tools, sub-agents
- Gateway — WhatsApp, Slack, Telegram, email, web, API
- Skills & Triggers — Scheduled automations and event-driven workflows
- API Reference — REST API and endpoints
- Self-hosting — Full deployment guide
- Changelog — What's shipped
Security
CASA Tier 2 Certified · TLS 1.3 in transit · AES-256 at rest · Your data is never used for model training · Self-host for full isolation
Security Policy → · Vulnerabilities: harshith@poozle.dev
Community
We're building the future of personal AI in the open. Come build with us.
- Discord — questions, ideas, show-and-tell
- CONTRIBUTING.md — how to set up and send a PR
good-first-issue— start here