Every AI agent demo eventually hits the same wall. The model can reason, draft, summarize and charm - and then you ask it to actually send the email, update the CRM, or refund the customer, and nothing happens. The gap between talking and doing is made of integrations: OAuth handshakes, token refresh, webhook signatures, rate limits, the unglamorous plumbing nobody puts in a keynote. Corsair, a two-person company out of New York and part of Y Combinator's Winter 2025 batch, has decided to own that plumbing so other builders don't have to.
The pitch is deliberately plain: "Connect your users to their apps." In practice, Corsair is an open-source TypeScript framework that sits between an AI agent and the outside world. You connect your Corsair instance to your agent once, and the agent immediately gets access to every integration you've configured - Gmail, Slack, Stripe, Notion, GitHub, HubSpot, Linear and roughly seventy others. Corsair handles authentication, refreshes tokens, verifies webhooks, isolates credentials per tenant, and enforces what the agent is and isn't allowed to do.
01 / THE PROBLEMThe integration tax
Ask anyone who has shipped a product that connects to other software, and they will describe the same tax. Every new integration is a small project of its own - read the API docs, wire up OAuth, store the tokens, handle the refresh, catch the rate limits, verify the webhooks, and then keep all of it working as the third party quietly changes things. Do it once and it is annoying. Do it for a product where each customer wants a different mix of apps, and it becomes a permanent maintenance burden that has little to do with the thing you actually set out to build.
Corsair's argument is that this work is the same everywhere, so it should be written once and shared. Instead of each team re-implementing a Slack connector or a Gmail sync, they import Corsair's and move on. For AI agents specifically, the stakes are higher: an agent that has to juggle raw credentials for a dozen services is both harder to build and riskier to run.
There is also a hidden cost that only shows up later. A connector is not a one-time build; it is a subscription to someone else's roadmap. APIs deprecate endpoints, rotate auth schemes, tighten rate limits and change webhook payloads on their own schedule, and every one of those changes lands as a bug in your product at an hour you did not choose. Centralizing the connectors means centralizing that maintenance too - when a provider shifts, it is fixed once for everyone rather than separately in a hundred codebases.
02 / HOW IT WORKSOne door, keys on the other side
The design detail that gives Corsair its character is what it does with credentials. Agents connect over the Model Context Protocol - the emerging standard for giving models tools - and when an agent wants to act, it doesn't receive an API key. It asks Corsair, which stores credentials with envelope encryption, resolves them internally, and executes the call on the agent's behalf. The agent gets the result, never the secret.
On top of that sits a permission system with four modes - open, cautious, strict and readonly - set per integration. And for the actions that can't be undone, Corsair adds a human in the loop. The canonical example from its own docs: ask an agent to email a file from Drive, and Corsair will retrieve the file and draft the message, then stop and create an approval request before it actually presses send. It is a small feature that says something larger about how the company thinks - the best automation knows when not to automate.
Multi-tenancy is where a lot of integration projects quietly fall apart, and it is a problem Corsair treats as a starting requirement rather than a later upgrade. Credentials and cached data are partitioned per tenant, so one customer's Slack token can never surface in another customer's session. For a solo script talking to your own accounts, that isolation is invisible. For a product built on top of agents, where every user brings their own set of connected apps, it is the line between a demo that works on stage and a system you can put in front of paying customers.
03 / THE PRODUCTThree ways to run it
Corsair comes in three shapes, aimed at three moods. If you want control, self-host the Apache-2.0 framework for free and run it on your own infrastructure. If you'd rather not run anything, point your agent at hub.corsair.dev and use a drop-in MCP URL. And if you want to provision connections, permissions and workflows programmatically, there's a managed cloud SDK documented at docs.corsair.dev. Underneath, it's a tidy TypeScript monorepo - pnpm workspaces, Turbo, Biome - which matters to the developers who are its first customers.
A sample of the plugin library. Each connector is open source - fork it, patch it, or write a new one.
04 / BUSINESS MODELOpen core, paid convenience
The money follows the classic open-core pattern: give the engine away, charge for the convenience of not operating it. The free Hobby tier covers 50 connections, 100,000 webhooks and up to three teammates. Pro, at $200 a month, unlocks unlimited connections and webhooks plus custom branding. Enterprise is priced case by case. The bet is that the first integration should cost nothing, and that by the time a company needs the hosted tier, Corsair has already become load-bearing.
It helps that the people deciding whether to adopt Corsair are also the people who can read its source. Its earliest users are developers and the startups building coding agents, and they tend to trust tools they can inspect. An open repository lets a skeptical engineer check exactly how credentials are stored before trusting the tool with production keys - a far easier sell than a closed platform asking for the same access on faith. The pricing meets that audience where it is: nothing to try, a flat fee when it becomes essential.
| Tier | Price | What you get |
|---|---|---|
| Hobby | Free | 50 connections, 100k webhooks, up to 3 members |
| Pro | $200/mo | Unlimited connections & webhooks, custom branding |
| Enterprise | Custom | Tailored limits, support and terms |
| Self-host | $0 | Apache-2.0, run it entirely on your own infra |
05 / THE MARKETThe agent-integration land grab
Corsair is not alone in noticing that agents need a way to reach the rest of software. It sits in a fast-filling category next to Composio, Paragon, Merge, Nango, Pipedream and, at the far consumer end, Zapier. What separates Corsair is the combination of choices: open source rather than closed, self-hostable rather than rented, agent-native rather than retrofitted, and built MCP-first rather than bolted on. In a market where the default is a proprietary connector catalog you pay to access, Corsair's Apache-2.0 license and public plugin library are the differentiator, not a footnote.
The ratio is the interesting part. Two people, seventy connectors, thousands of stars. That math only works if the community writes plugins alongside the founders - which is precisely what an open license is designed to make happen. YC partner Gustaf Alstromer has publicly noted more startups and coding agents picking Corsair as their integration layer, a useful early signal for a project this young.
What people actually do with it is straightforward once the plumbing disappears. A team building a customer-support agent wires it to Gmail, Slack and Linear and lets it triage, reply and file tickets. A finance tool connects Stripe and Google Sheets to reconcile payments on request. An internal assistant reaches into Notion, HubSpot and Calendar to answer questions and schedule follow-ups. In each case the interesting work is the agent's judgment, not the twelve connectors underneath it - and hiding those connectors is exactly the service Corsair sells. The company frames the whole thing as letting builders construct custom agents, dashboards and workflows across hundreds of integrations without owning the machinery that makes them talk.
06 / THE FOUNDERFrom Demeter to Corsair
Corsair is led by Dev Jain, who studied mechanical engineering and economics at UT Austin and wrote software at Curri (YC S19) before founding the company in 2024. Corsair was not his first swing in the same batch: he and a co-founder first built Demeter, an AI back office for private investors, whose portfolio-management tool automated performance tracking and return analysis. The old handle - the LinkedIn page still reads "joindemeter" - is a fossil of that earlier idea.
The pivot reads less like a retreat than a widening of aim. Building an automation product for investors meant wiring up the same integrations Corsair now generalizes; the connective tissue turned out to be a bigger problem than the vertical it was serving. Sometimes the tool you build to finish a job becomes the job.
That origin also shapes what Corsair is good at. The expertise on display is not a novel model or a flashy interface - it is the patient, defensive engineering of systems that must survive contact with other people's APIs. Envelope encryption for stored secrets, signature-verified webhook handlers behind a single endpoint, typed connectors, permission modes and approval gates: these are the concerns of a team that has been burned by integrations before and is building the thing they wished they'd had. It is unglamorous work, and that is rather the point.
Whether Corsair becomes the default layer or one option among several will depend on the unglamorous things: how fast the plugin library grows, how reliable the hosted hub proves, and whether the open-source goodwill converts into paying teams. For now, it is doing the specific, useful thing of turning "connect your users to their apps" from a quarter of engineering work into an import statement - and keeping your keys to itself while it does.