redplanethq/core

TypeScript1,404129187 issues16 contributorsNOASSERTION

Summary

CORE is a self-hostable 'digital brain' platform that sits as a persistent memory and action layer between users and their AI tools (Claude, Cursor, Windsurf, etc.) via MCP. It stores user context in a Neo4j temporal knowledge graph, exposes 200+ actions across 50+ apps (Gmail, Slack, Linear, GitHub, etc.) through a unified MCP endpoint, and includes an orchestrator agent that can proactively trigger workflows. The webapp is a Remix/TypeScript app backed by Postgres (Prisma) and Neo4j, with integrations as standalone TypeScript packages.

Great for

building persistent cross-tool memory and action layers for AI assistants — specifically if you're interested in how LLM agents can maintain user context across sessions and orchestrate multi-step workflows across SaaS apps via MCP

Easy wins

+Add a new integration: there's a documented template at docs/opensource/building-integrations.mdx and a GitHub issue template for it — the existing Gmail/Slack/Linear integrations follow a clear pattern (Zod schemas + getTools() + callTool() switch) that you can copy directly
+Fix one of the 95 'good first issue' labeled issues — many are tagged 'integration', 'ui', or 'documentation' which are self-contained
+Write a CONTRIBUTING.md: the repo has no contribution guide at all, which is a real gap given 187 open issues and 95 good-first-issues — this alone would be a high-impact PR
+Add error handling to integration MCP handlers: the Linear/Slack/Gmail integrations have try/catch at the top level but individual tool branches (especially in the large switch statements) often throw raw errors without structured MCP error responses

Red flags

!Zero test coverage — not a single test file found anywhere in the repo, including for the knowledge graph operations which are the core value proposition
!License is listed as NOASSERTION — unclear licensing situation that could be a blocker for enterprise contributors or forks
!commit_frequency_30d shows 1 commit in the last 30 days despite 187 open issues — activity metrics suggest this may be primarily maintainer-driven with low external contribution velocity
!The _addComparisonToResult function in github-analytics/src/analytics.ts is dead code with a misleading JSDoc comment — minor but indicates some code drift
!No secrets scanning or security workflow in .github/workflows — OAuth tokens are passed around as plain strings through config objects, and the webapp handles OAuth credentials; no evidence of rotation or encryption at rest

Code quality

decent

The integration packages (Gmail, Slack, Linear) are well-structured with consistent patterns: Zod schemas → zodToJsonSchema → MCP tool definitions → switch-based dispatch. The persona generation code (aspect-persona-generation.ts) is surprisingly sophisticated with thoughtful comments explaining design decisions (e.g., why BERT/HDBSCAN was dropped). However, there are no tests anywhere, error handling in integration callTool() implementations often just throws raw Error objects rather than returning structured MCP errors, and the GitHub Analytics code has a dead helper function (_addComparisonToResult) that's explicitly marked 'Reserved for future use' but never called. The Linear integration's callTool() switch statement is several hundred lines of copy-pasted GraphQL with no abstraction.

What makes it unique

This is genuinely differentiated from the crowded 'AI memory' space (Mem0, Zep, etc.) by combining memory *and* a unified MCP action layer in a self-hostable package with a real web UI. Most competitors are pure libraries or cloud-only. The temporal knowledge graph approach (classifying facts by aspect: Preference, Directive, Goal, etc.) is a more structured take than raw vector search. That said, the 'digital brain' category is getting crowded fast, and the MCP ecosystem is evolving rapidly enough that some integration glue code may become obsolete.