// the find
airweave-ai/airweave
Open-source context retrieval layer for AI agents
Airweave is a self-hostable data sync and retrieval layer that connects 50+ SaaS sources (Slack, Notion, Jira, Salesforce, etc.), keeps them indexed, and exposes a unified search API for AI agents to query. It sits between your existing tools and your agent stack, handling auth, incremental sync, and vector search so you don't wire up a separate pipeline per integration. Target is teams building agents that need to ground responses in internal company data.
The adapter pattern throughout the backend is genuinely disciplined — every external dependency (LLM, cache, email, identity, reranker) has a protocol interface plus a fake implementation, which makes unit testing tractable without mocking infrastructure. Temporal for sync orchestration is the right call: you get durable execution, retries, and visibility for free, instead of a fragile cron-plus-database-queue setup. MCP server support means agents using Claude, Cursor, or similar tools can connect without SDK glue code. The breadth of integrations (50+) is real — not stub connectors, each has its own cursor-based incremental sync implementation.
The self-hosted footprint is punishing: you're spinning up PostgreSQL, Vespa, Redis, Temporal, and the backend+frontend together. Vespa alone wants 6+ GB of RAM and is a JVM distributed search engine — most self-hosters will fight it before they fight anything else. The identity layer defaults to Auth0, so running without it means using the null/fake provider, which works but leaves a gap in the docs around what that actually means for production use. Writing a new connector requires understanding entity definitions, cursor architecture, and the integration YAML format — the .cursor/rules files hint that this is genuinely complex, not just 'implement an interface'. Agentic search silently degrades if no LLM key is configured; the fallback behavior isn't documented clearly enough for someone standing this up for the first time.