Connectors to a dozen SaaS apps. OAuth flows for each one. Schema detection that survives the next API change. Document chunking across DOCX, PDF, PPTX, XLSX. An embedding pipeline. A vector store. A search API. ACL enforcement so agent queries only return what the end user is allowed to see. Incremental sync so the index doesn't go stale by Friday. None of it is novel — but every agent team writes a half-finished version of it themselves, and the half-finished version is exactly what breaks the first time a customer's Slack has 200,000 messages.

The OpenAI MCP launch made the problem explicit. Every remote MCP server is now expected to implement search and fetch — to behave like a search engine over a customer's data.^[6] That's the entire retrieval stack as a public contract: ingestion + indexing + ranking + access control. Every team adopting MCP either rebuilds it or rents it. Airweave is the rent option, open-sourced.

The cost isn't writing the code once. It's the long tail — keeping forty connectors green as their APIs drift, keeping the index fresh as customer data changes, keeping permissions correct as agents query across workspaces. That's not a hackathon project; it's a company. And every agent founder we talked to would rather not own it.

MCP turned retrieval into a protocol. OpenAI's remote MCP launch (Jun 4 2025) mandated search and fetch on every server — a public spec for what agents need from data sources.^[6] Anthropic's MCP spec did the same on the client side.^[10] What used to be a thousand bespoke integrations is now a single contract — and a single contract is exactly the surface a default platform can fill.

The connector matrix is being rebuilt everywhere, badly. Every agent startup we talked to has a half-finished integrations folder; every one of them would rather not own it. The work is repetitive, the maintenance is endless, and the differentiation is zero — the perfect shape for an OSS default to absorb. Composio is already a paying Airweave customer; Cursor and Claude Desktop are design partners.^[3] The category is forming in real time.

Frontier labs are keeping the plumbing closed. ChatGPT Connectors, Claude's data integrations, Codex's VM — all internal-only.^[6] As OpenAI and Anthropic evolve from model vendors to product companies, holding the retrieval middleware private preserves their UX edge. That creates the structural opening for an open, vendor-agnostic stack that every agent builder outside the walled gardens can adopt. The same dynamic that gave Vercel space against AWS and Supabase space against Firebase.

A year on, the protocol bet resolved. PulseMCP alone lists 17,974 MCP servers as of June 2026^[20]; tracked counts across the major public registries grew roughly 38% in the four months to mid-April 2026 (~6,800 → ~9,400).^[21] Every one of those servers is expected to behave like a search engine — and most of their authors don't want to build the retrieval stack behind it. Capital confirmed the read: Mem0 raised $24M for agent memory with AWS naming it the exclusive memory provider for its Agent SDK^[17], Letta took $10M from Felicis^[18], and Glean re-rated to $7.2B on the enterprise side.^[19] The context layer is no longer a thesis; it's a funded category — and the open, source-of-record retrieval slot in it is the one Airweave occupies.

One API. Every agent builder wants the same thing — a single search endpoint that returns the right rows from across a customer's tools. Airweave's contract is that endpoint, with the connector graph behind it and the ACL filter at query time.^[1]

Two surfaces. REST for traditional apps; MCP server for agents. Same data plane, two interfaces — and the MCP interface is exactly the shape OpenAI's docs now mandate.^[6] Builders adopting MCP this year inherit Airweave's retrieval for free.

One permission model. Every connector replicates the source app's ACL into Airweave's index. Queries filter by the end user's source-app permissions in real time. No accidental cross-tenant leakage, no agent returning rows the user can't see in Slack itself. The unsexy work that becomes the customer's procurement requirement at deal #20.

The compounding work isn't novel — that's the point. OAuth handshakes for forty SaaS apps. Schema detection that survives the next API change. Incremental sync that doesn't lose deletes. Rate-limit handling that doesn't take down the customer's workspace. ACL replication so agent queries return what the end user is allowed to see. None of it is glamorous; all of it has to work.^[3]

Switching costs compound with every connector. Once a customer is wired into your connector graph, swapping you out means re-doing OAuth for every source app, re-validating every permission model, and re-running an initial sync that might take hours. The cost compounds with every connector added, every customer onboarded, every workspace authenticated. That's the same shape that made Plaid and Merge durable — and the same shape Airweave is building one OSS pull request at a time.

OSS is the accelerator, not the threat. Each connector is a scoped contribution — exactly the unit of work the community is good at, and Airweave is already taking PRs against it. The open license recruits the long-tail integrations the team would never prioritize in a closed roadmap. The hosted plane runs everything reliably and absorbs the operational cost.

The agentic AI tools market is forecast to grow from $6.2B in 2024 to $419B by 2034 at ~52% CAGR.^[5] Retrieval is the unglamorous half of that — every agent in production needs to read customer data, and every team would rather buy than build the connector matrix. The bet is that the unified context layer for agents becomes infrastructure on the same scale as the database was for web apps.

Inside YC alone, Composio is already a paying Airweave customer; design partners include teams at Cursor and Claude Desktop.^{[3] [4]} Beachhead math: ~50k AI dev teams × $200–$2,000/mo on the cloud tier ⇒ $120M–$1.2B ARR in the near-term ICP before Airweave touches a single F500.

The demand side is no longer hypothetical. Mem0 — the adjacent memory layer, same buyer — reported API calls growing from 35M in Q1 2025 to 186M in Q3 2025, a 5.3x jump in two quarters.^[17] That's the consumption curve of agent-context infrastructure being wired into production, and retrieval over systems of record sits one layer below memory in the same stack.