CORE Personal AI OS
An always-on personal AI assistant with persistent memory, task execution, and multi-interface access — open source and self-hosted.
Evidence: README claims 'human-in-loop by default' and 'structured task planning with human approval' but lacks implementation details; SECURITY.md mentions 'principle of least privilege' and RBAC but no concrete configuration; data flow transparency only mentions TLS 1.3 and AES-256 without specifics; sensitive data handling only mentions token encryption; dependency security only mentions updating dependencies without audit or lockfile; external effects only mention browser isolation; rollback not mentioned; source attribution only mentions Poozle Inc. and copyright without contributors or verification. Deductions: lack of concrete evidence, mostly assertions.
Evidence: README and SECURITY.md have version inconsistency (README doesn't mention version, SECURITY.md supports 0.1.x, package.json is 0.7.20), but overall description is consistent; dependency availability only mentions Docker and Node requirements without dependency list; failure messages not mentioned. Deductions: version inconsistency and lack of failure handling documentation.
Evidence: README describes multiple usage scenarios (voice, messaging, browser) but doesn't specify target audience; capability boundaries only mention '1000+ actions' and '50+ apps' without limitations; trigger precision only mentions 'within 3 minutes' without trigger conditions; environment fit mentions Docker and Railway but not other environments. Deductions: unclear boundaries and trigger conditions.
Evidence: README has clear structure with quickstart, docs links, security section; install notes provide npm and Railway deployment; naming stable but version inconsistent; examples and FAQ have examples but no FAQ; known limitations not mentioned; license is AGPL-3.0 with Commons Clause but metadata is NOASSERTION; versioning changelog has link but no content; maintenance responsibility only mentions community and Discord. Deductions: version inconsistency, no FAQ, no known limitations, license metadata mismatch.
Evidence: Output usability has screenshots and demo but no actual output examples; marginal value has comparison table but no independent verification; cost benefit mentions 'no per-token bill' but no cost analysis. Deductions: lack of actual outputs and cost data.
Evidence: README claims 88.24% benchmark accuracy but no reproduction method; cross-source verification only mentions benchmark repo without other sources; fact vs inference separation not clear. Deductions: lack of verifiable details.
- Version inconsistency: README doesn't mention version, SECURITY.md supports 0.1.x, package.json is 0.7.20, which may affect user judgment on supported versions.
- License metadata is NOASSERTION, but LICENSE file is AGPL-3.0 with Commons Clause, which may cause legal risks.
- Security claims are mostly assertions without concrete implementation evidence, such as least privilege and RBAC specifics.
- Benchmark claims 88.24% accuracy but no reproduction method, cannot be independently verified.
What does this agent do, and when should you use it?
CORE is an open-source, self-hostable personal AI operating system that goes beyond a chatbot: it runs continuously in the background, watches your digital life, and proactively handles tasks. You can interact through a Scratchpad (daily page), voice, messaging apps (WhatsApp, Slack, Telegram), and a web interface. It maintains a temporal knowledge graph of your emails, meetings, GitHub, Linear, Slack, and conversations, so every task starts with full context. CORE can delegate coding work to agents like Claude Code and Codex, run them locally or in Docker/Railway, and open pull requests when done. It integrates with 50+ apps via MCP, supports webhook triggers, and defaults to human-in-the-loop approval. Installation is via npm package @redplanethq/corebrain, and data stays in your own infrastructure.
CORE indexes email, meetings, GitHub, Linear, Slack, and every assistant conversation into a temporal knowledge graph. Users write tasks in the Scratchpad (at http://localhost:3033), e.g., '[ ] Fix the auth bug from issue #47', and CORE picks it up within 3 minutes, gathers context, drafts a plan for approval, then executes it. Tasks can be one-shot or recurring, each with a dedicated chat thread, and can spawn coding sessions (Claude Code, Codex), browser automation, and terminal commands. Voice input is available via Ctrl+Option on Mac; messaging via WhatsApp, Slack, or Telegram. After setting up the gateway with corebrain gateway setup, CORE can run coding agents, drive a browser, and access local folders.
- A developer writes down a task on their daily Scratchpad, and CORE automatically pulls context from GitHub and memory to fix an issue.
- A product manager sends a task from the airport via WhatsApp, and CORE executes it with the same memory and context as if it were initiated locally.
- A user delegates a coding task to CORE, which runs a Claude Code session in the background, opens a PR, and continues even when the laptop lid is closed.
- A team connects Slack to CORE, uses 50+ MCP integrations (GitHub, Linear, Jira), and receives proactive work from webhook events.
- An individual customizes CORE's name, personality, and voice, then uses voice commands to capture tasks without breaking flow.
What are this agent's strengths and limitations?
- Persistent memory: temporal knowledge graph gives every task full context, avoiding repeated explanations.
- Multi-interface access: Scratchpad, voice, messaging, and web let you use CORE however you work.
- Leverages existing subscriptions: can run on your Claude Pro or Codex subscription without extra API keys.
- Self-hosted and open source: you retain full control over data, and it's never used for training.
- Extensive integrations: 50+ app connectors via MCP, with webhook triggers for proactive automation.
- High installation requirements: needs Docker and 4 vCPU/8GB RAM, which can be a barrier for resource-constrained users.
- Dependency on third-party subscriptions: core execution depends on Claude Code or Codex, creating provider lock-in.
- Learning curve: requires understanding concepts like Scratchpad, Tasks, Gateway; configuration is not trivial.
- License ambiguity: GitHub shows 'NOASSERTION' despite README stating AGPL 3.0, which may pose legal uncertainty.
- Benchmark details are not fully public: only reports 88.24% on LoCoMo without complete baselines in the repo.
How do you install or deploy this agent?
Requirements: Docker 20.10+, Docker Compose 2.20+, 4 vCPU / 8GB RAM. Run npm install -g @redplanethq/corebrain && corebrain setup and follow the wizard to specify install directory, AI provider, API key, and chat model. The installer generates secrets, starts the stack, and opens http://localhost:3033. Alternatively, one-click deploy to Railway.
How do you use this agent?
After installation, run corebrain login and corebrain gateway setup to connect a gateway for coding agents, browser, and local folder access. Open the Scratchpad, type a task like '[ ] Summarize my open GitHub issues', and approve the plan CORE proposes. It executes and returns results. You can also configure messaging integrations (WhatsApp, Slack, Telegram) and voice input.
How does this agent compare with similar options?
Compared to OpenClaw, CORE offers multiple interfaces, persistent memory, and structured task planning. Compared to Hermes Agent, CORE adds voice and messaging interfaces. Compared to Devin/Copilot, CORE is open source and self-hostable, with more interface options.