The two-person, Y Combinator-backed team built a docx MCP - plumbing that turns raw agent output into tracked, review-ready .docx files. It is a quiet bet on the least glamorous corner of the AI stack.
Every founder story in AI right now sounds the same: an agent that reasons, plans, and acts. Vespper's founders lived inside that story and noticed the part nobody demos - the moment the agent finishes thinking and a human still has to copy the answer out of a chat window, paste it into Microsoft Word, fix the formatting, and turn on tracked changes before anyone in legal will look at it. That unglamorous handoff is now the whole company.
Vespper (YC F24) describes its product in six words: the docx MCP for agents. Translated, it is a piece of infrastructure that lets an AI agent read and edit Microsoft Word documents through the Model Context Protocol - without the agent, or the developer, ever touching a line of Office Open XML. Connect an agent, and Vespper turns its raw output into a clean, formatted, tracked-changes, review-ready .docx file.
It is a narrow product with a specific audience, and that is the point.
"The docx MCP for agents - connect your AI agents via MCP to turn raw output into clean, tracked, review-ready Word documents."
Vespper, company descriptionHere is the trap that catches a lot of enterprise AI. The model produces something genuinely good - a contract summary, a regulatory filing draft, a compliance memo. In the demo, everyone claps. Then a real customer asks the quiet question: can I get that as a Word document, in our template, with tracked changes, so my reviewers can mark it up? And the answer, for most agent setups, is no. Not cleanly.
In regulated industries - law, pharmaceuticals, finance - the finished artifact is not a chat message. It is a .docx file with a redline history that an auditor can follow and a human can sign off on. The document is the deliverable. An agent that stops at plain text has done the interesting 90% and skipped the 10% that decides whether anyone can use the work.
Vespper is not trying to be the agent. It sits underneath whatever agent a team already runs, exposed as a remote MCP endpoint. The Model Context Protocol has become a common language for connecting agents to tools and data, and Vespper speaks it natively. An agent framework points at Vespper, and Vespper handles the tedious machinery of building a valid Word file: styles, numbering, tracked changes, the structural rules that make a .docx open correctly instead of throwing a repair prompt.
The reason this is worth a company is that .docx is genuinely miserable to generate by hand. It is a zipped bundle of XML with decades of accumulated rules. Nobody wakes up wanting to write that code. Vespper's wager is that a lot of teams will happily pay to never learn how.
Vespper did not start here. Its first product was an open-source, 24/7 AI on-call engineer - a multi-agent system that lived in Slack, listened for production alerts, investigated them, and posted root-cause hypotheses within seconds. It plugged into the observability and incident stack teams already ran: Datadog, Grafana, Prometheus, PagerDuty, Opsgenie, plus GitHub, Notion, Jira, and Confluence for context. It was self-hostable and shipped under the Apache-2.0 license, so teams could own their data.
The company launched it on Launch YC in late 2024. Then it did the thing that is easy to describe and hard to do: it changed direction.
"Vespper is an AI on-call engineer running 24/7 to troubleshoot your alerts and surface the right data so incidents never fall through the cracks."
Vespper's original Launch YC descriptionThe through-line between the two products is clearer than it first looks. Both start from a repetitive, thankless task that a smart person resents doing - being paged at 3am, or hand-formatting a document for the tenth time - and hand it to a machine. The pivot changed the address, not the instinct.
Topaz Turkenitz, co-founder and CEO, came from Snyk, the security unicorn, where she worked on distributed systems and led growth experiments - including a project she has said helped land a Fortune 500 bank. Her background blends computer science and design. Her co-founder and CTO, David "Dudu" Lasry, spent years in machine learning and full-stack engineering at companies including Viz.ai, the medical-AI unicorn, and Google.
Between them they have spent a lot of nights being woken by systems that could not finish a job on their own. That is a useful thing for a founder to have felt. It tends to produce products aimed at real chores rather than imagined ones. The company also lists an advisor with a background in regulated R&D workflows at large pharmaceutical firms - a tell about which industries Vespper thinks it fits.
The market is crowded with agents and thin on the connective tissue those agents need. Vespper is not alone in the specific niche - there is a small field of Word-oriented MCP servers, including open-source projects like GongRzhe's Office-Word-MCP-Server and SecurityRonin's docx-mcp, plus Office-365 tool suites from companies like Arcade. What Vespper is betting on is focus: a hosted, managed endpoint pointed squarely at teams whose output has to be a review-ready Word document, sold to the people building agents rather than to the people who never wanted to think about file formats.
Microsoft Word is 40 years old and it is not going anywhere. The contrarian move is not to replace the document. It is to be the bridge to it.
The Vespper thesis, in one lineThere is a version of the next decade where the flashy agent companies commoditize each other and the durable margins sit in the boring layers - the connectors, the exporters, the format handlers nobody wanted to build. Vespper is a small, early, unproven bet on that version. Whether it is right will come down to how many teams decide that generating a Word file is a problem worth outsourcing, and whether Vespper's hosted service stays enough ahead of the free open-source alternatives to charge for.
Reported funding is modest and early - a seed-stage figure in the low hundreds of thousands, alongside the Y Combinator investment. The team is two people in Tel Aviv. Read that as a company still finding the shape of its market, not one that has already won it. The interesting part is the discipline: instead of adding features to a product that was not landing, they cut to a smaller, sharper problem.
Note: Public figures on Vespper's headquarters and funding vary across sources - some list San Francisco, some Tel Aviv; reported seed funding ranges from roughly $130K to $500K. Figures here reflect a range of public sources as of July 2026 and are approximate. No official product demo video was found at publication; check Vespper's website and GitHub for the latest.