CopilotKit Lands $27M, Crowds Out Vercel's AI SDK

Half the Fortune 500 already builds on a protocol most engineering leads could not name on Friday. the lead, not the $27 million. CopilotKit closed its Series A on Tuesday with Glilot Capital, NFX, and SignalFire writing the checks, putting fresh capital behind a Seattle twenty-person team that became the de-facto AG-UI standards body while the framework camps were still arguing about hooks. The question for an engineering lead picking an agent SDK this quarter changed under them, and most have not realized it yet.

The Deployment

CopilotKit, originally incorporated as Tawkit Inc., raised $20 million in Series A capital plus $7 million in a previously unannounced seed round. Glilot Capital led, with NFX and SignalFire following. The Seattle company has roughly twenty employees, most of its engineering staff in-market, and plans to put the round into growing that local headcount.

The product the round is funding sits in a specific slot of the AI-agent protocol stack. AG-UI, short for Agent-User Interaction, is the open standard CopilotKit shipped for how agents talk to humans inside an existing software application: generating interactive charts on the fly, updating dashboards, and taking actions through the app's interface. Google, Microsoft, Amazon, and Oracle have adopted it. CopilotKit says more than half of the Fortune 500 use its tools, primarily as users of the open-source project, with named enterprise customers including Deutsche Telekom, Docusign, Cisco, and S&P Global running the paid CopilotKit Enterprise Intelligence tier.

The open-source core sits at over 40,000 GitHub stars with millions of installs per week. The paid product self-hosts and layers on persistent conversation threads, analytics, and real-time learning. The competitive set, as CopilotKit positions it, is Vercel's AI SDK, Assistant-ui, and OpenAI's Apps SDK. CopilotKit's pitch against all three is horizontal and vendor-neutral: it works with whichever agent framework, cloud, and backend a company already runs.

The two co-founders are brothers. Atai Barkai, the CEO, came out of Meta media infrastructure and led iOS app development at Doximity before this; he holds physics degrees from Penn. Uli Barkai runs growth and partnerships, with finance and philosophy backgrounds from Columbia and Tel Aviv. They originally co-founded the company as a podcast platform, open-sourced their internal copilot tooling, and pivoted when developer interest outpaced the original product. The 2023 Techstars Seattle cohort was the on-ramp.

Why It Matters

The agent-protocol stack is doing in 2026 what the streaming-data stack did in 2017: settling into one default per slot. Agents need three connections to be useful in production. They need to call tools and read context: MCP, the Model Context Protocol. They need to talk to other agents and compose workflows: A2A, Agent-to-Agent. And they need to render results and accept input from a human inside the application: AG-UI. Three slots, three protocols, and the cycle is rewarding the project that ships the cleanest contract first.

When the same four cloud providers (Google, Microsoft, Amazon, Oracle) adopt the same protocol in the same year, the result is not polite endorsement; it is a buyer-side moat compounding. The framework alternatives now have to either ship interop or watch their default status erode every quarter. Vercel's AI SDK has the developer-experience polish but no comparable cloud-side endorsement on the protocol layer. OpenAI's Apps SDK has reach but is structurally vendor-aligned, which cuts off exactly the horizontal customer that CopilotKit closes most easily. Assistant-ui is closer architecturally and harder to dislodge in greenfield projects, but it does not have the Fortune 500 install base to point at in a procurement meeting.

The pattern this echoes most directly is open-source infrastructure plays from the prior decade. Confluent did it with Kafka. HashiCorp did it with Terraform. MongoDB did it with the document model. Open core, broad free distribution, paid enterprise tier on a self-hosted base, multi-cloud customer logos on the homepage. Procurement already has a template for that contract; legal has reviewed the license; security knows what self-hosted means in their network diagram. CopilotKit walking the same path with a twenty-person team is unusual but not unprecedented. Confluent had similar headcount when its early enterprise customers came on board.

What Other Businesses Can Learn

If you run an engineering team at a 50-to-500-person firm and you have any agent project on the roadmap for this fiscal, treat the AG-UI question as architectural before you treat the SDK question as ergonomic. The framework you pick is locked into a protocol whether you noticed or not. Audit that first.

Pin the protocol layer before the framework. Standardize on AG-UI, MCP, and A2A as the three contracts your internal agents speak. The framework underneath becomes interchangeable, and the framework vendors have to compete on developer ergonomics rather than lock-in. That puts you in a stronger procurement position when contract renewal comes around eighteen months from now.

Treat the open-source core as production infrastructure, not a community side project. Forty thousand GitHub stars and millions of weekly installs put CopilotKit in the same risk profile as a HashiCorp Terraform provider: wide enough that breaking changes get flagged loudly, narrow enough that you can read the issue tracker and predict the next release. Pin your AG-UI dependency, watch the changelog on minor bumps, and assume the spec stabilizes faster than your internal agents do.

Use the named enterprise customers as your reference list. Deutsche Telekom, Docusign, Cisco, and S&P Global cleared procurement at four very different companies in four very different industries. not a marketing list. It is a legal-and-compliance pre-clearance you can hand to your own counsel and shorten the review cycle by weeks.

If the cloud you already pay is on the same protocol as the SDK you are evaluating, the renewal cycle is what selects the winner now, not the developer-experience checklist.

The cost of getting this wrong is not catastrophic. Agent code is small enough that a year-from-now migration is real but bounded. The cost of getting it right is that the protocol decision compounds. Every internal agent you ship under AG-UI is one more service your team writes against a contract the cloud providers have already endorsed, and each one shortens the next procurement conversation.

Looking Ahead

Watch for AWS to formalize AG-UI support as a Bedrock primitive rather than a third-party adoption note. the moment the framework competitors' interop catch-up has to ship or stop being serious. The other signal worth tracking: whether Vercel adds AG-UI to the AI SDK's exported surface in the next two minor versions, or whether they build a parallel protocol. Pick the second and they fragment the stack; pick the first and they concede the standards-body slot to a Seattle twenty-person team.

Sources

• Seattle's CopilotKit raises $27M, as some of the biggest names in tech adopt its AI agent protocol, accessed 2026-05-05