The scene is familiar to anyone who's wrestled with code past midnight: your automated assistant, whichever flavor you're running, has hit a wall. Maybe it's looping on the same React bug. Maybe the deployment config refuses to cooperate. In the old workflow, you'd ping a colleague on Slack or file a ticket and wait. In the new one emerging from a startup called Humwork, the AI does something else entirely—it asks for help itself.
No human intermediary required. The agent reaches out, gets matched with a vetted expert in under half a minute, trades context, and resumes its task with a fix in hand.
That, at least, is the promise from Humwork, which emerged from Y Combinator's Spring 2026 cohort with a singular pitch: the first real-time marketplace where autonomous agents escalate to humans when they're stuck. The mechanism underpinning it all? Anthropic's Model Context Protocol, an open standard that's become something of the hidden plumbing in the world of agentic AI.
Whether this marks a genuine inflection point in how software gets built—or just a clever MCP server with a very specific use case—depends on questions the company hasn't fully answered yet.
The Protocol Beneath It All
Humwork's entire architecture rests on MCP, the JSON-RPC-based protocol Anthropic unveiled in late 2024 to standardize how AI agents interact with tools and external data sources. By early 2026, momentum had already built: over 10,000 active MCP servers and 97 million monthly SDK downloads were documented in the ecosystem. The ecosystem spawned curated registries, enterprise plays like Workato's MCP platform, and a growing list of security papers documenting the usual suspects—prompt injection, tool poisoning, misconfiguration risks.
For Humwork, MCP is both the trigger and the delivery system. When a compatible agent—Claude Code, Cursor, Lovable, Replit, among others—loops on a problem or encounters ambiguity, it can invoke Humwork's MCP server with a single call. Setup, the company claims, takes around 60 seconds: install the server, and the agent gains access to what Humwork's homepage describes as "3,000+ verified experts" standing by around the clock.
The handoff itself is straightforward, almost deceptively so. An agent stuck on something—debugging a deployment issue, refining marketing copy, untangling a legal clause—calls Humwork. The platform matches a domain expert in less than 30 seconds. That expert sees the full context: code snippets, error logs, prior attempts, all with personally identifiable information automatically scrubbed. The conversation happens directly over the MCP channel, and the solution flows back into the agent's working memory.
Humwork's site, as of early May 2026, advertises an 83% resolution rate and an average first reply under two minutes. Earlier reports from mid-April cited an 87% resolution rate across 2,858 questions during beta, with a then-smaller pool of "1,000+ experts." The discrepancy suggests rapid scaling. Or rapid messaging. All figures are self-reported; none have been independently audited.
Two Founders, One Big Bet
Behind Humwork are co-founders Yash Goenka and Rohan Datta, both UC Berkeley graduates with prior stints in AI infrastructure. Goenka, still an undergraduate when some of his earlier projects launched, holds a patent related to graphene supercapacitor manufacturing and built his first large language model startup back in 2021. His portfolio includes ventures like phonecall.bot and Jarvys. Datta earned both a bachelor's and master's from Berkeley, worked as a data scientist at MicroGrid Labs, and previously built an AI voice-calling platform that automated north of one million minutes of calls.
The duo operates lean—perhaps leaner than you'd expect for a Y Combinator company with enterprise ambitions. LinkedIn data from mid-April pegged headcount at "2–10 employees," while the company's YC profile lists the team size as 2. Humwork's YC profile lists open go-to-market roles, complete with compensation and equity ranges, as of early May. Tyler Bosmeny is the company's primary YC partner.
It's a team of two betting that the future of work isn't just AI agents doing more tasks, but AI agents knowing when to stop and ask a human.
What It Costs (If You Can Find Out)
Humwork runs on a pay-per-use escalation model. Developers and teams pay when their agents invoke human help; experts get compensated per session. The company's terms of service mention a credit system and tiered plans, but the actual checkout pages remain gated. No public price sheet.
India Today, citing correspondence with Humwork support in mid-April, reported two consumer plans priced at $20 and $100, enterprise options, and a flat expert payout of $0.70 per minute during active chat sessions. The outlet Inshorts echoed that 70-cent figure. Neither number appears on Humwork's public site as of early May, and the company hasn't confirmed them independently.
The domains Humwork covers span software engineering, design and UX, marketing and copywriting, product strategy, legal, and finance. Experts are vetted through identity verification, skills assessments, and domain-specific testing, according to public materials. What "vetted" means in practice—whether it's a ten-minute screen or a multi-round technical gauntlet—remains unclear.
The Funding Picture
Humwork went through Y Combinator's Spring 2026 batch. Multiple aggregator platforms suggest the company raised $500,000, and LinkedIn's company page lists a pre-seed round of that amount dated October 9, 2025, with YC as the investor. CB Insights, in a late April crawl, noted $500,000 raised via convertible note roughly 22 days prior to the snapshot. These figures appear across multiple sources but lack primary filing verification. The specifics—valuation, dilution, any follow-on interest—remain unconfirmed.
The company launched publicly around April 15, 2026, with a YC Launch post that triggered a wave of coverage. Analytics Drift, The Agent Times, TestingCatalog, and India Today all published features within days, latching onto the conceptual inversion: AI as client, humans as service layer. Y Combinator amplified the launch on LinkedIn mid-April, and tech commentators framed it as an early marker in the so-called "agent economy."
Whether that economy is real or still mostly theoretical is another question.
A Crowded Field, Suddenly
Humwork isn't the only player chasing the MCP-to-human handoff. HumanMCP.ai offers AI-to-human task delegation for phone calls, research, and verification tasks. Human-mcp.io bills itself as "a marketplace where AI agents hire humans via MCP." RentAPerson.ai and Expert Sapiens both expose MCP servers or REST APIs designed for agent-driven expert access. Tendem.ai claims a pool of "10,000 experts" and promotes its MCP integration.
Then there's Agentalent.ai, launched in late March 2026 by Monday Agent Labs—a marketplace for hiring AI agents, not humans. An agent-to-agent play, sure, but a signal that the category is heating up.
The broader MCP ecosystem is maturing quickly, perhaps more quickly than its security posture. Multiple papers published between 2025 and early 2026 have documented risks including prompt injection, misconfiguration, and tool poisoning. Humwork mentions PII redaction and expert vetting in its materials but doesn't detail specific countermeasures for these threat vectors. That might be fine for early adopters. It might not be fine for enterprises with compliance obligations.
What's in the Fine Print
Humwork's privacy policy, effective April 8, 2026, outlines data handling for clients, experts, and app interactions. Voice-based qualification assessments—part of the expert onboarding—may be transcribed by third-party speech-to-text services in real time. Expert-provided content, meanwhile, may be used for AI R&D and dataset licensing. That provision underscores something worth noting: Humwork isn't just a marketplace. It's also a data flywheel, capturing high-quality human responses to agent failures and likely feeding that back into model training or resale.
The terms of service include the standard arbitration clauses, class action waivers, and credit-based payment structures. Third parties may process audio in voice interactions, which means the "human expert" layer introduces its own compliance surface. For a product targeting enterprise deployments, that's not a small detail.
The Questions That Remain
As of early May 2026, Humwork hasn't publicly named a single enterprise customer. The homepage invites companies deploying AI agents at scale to "book a call," but there are no logos, no case studies, no testimonials from recognizable names. The supported platforms are listed—Claude Code, Cursor, Codex, Lovable, Replit, OpenClaw, ChatGPT, Claude, Gemini, Cowork—but real-world integration examples remain scarce.
The expert pool is growing, or so the company says. Whether it's 1,000, 3,000, or somewhere in between is hard to pin down from the outside. More importantly, the quality, breadth, and actual availability of that pool under sustained load remains opaque. An 83% resolution rate sounds impressive until you ask what happens to the other 17%. Does the agent loop indefinitely? Does the user get pulled in after all?
And then there's the core assumption: that AI agents will increasingly operate autonomously, and that the right safety valve isn't better models or more compute, but a 30-second line to a human who knows the answer. If MCP adoption continues its early 2026 trajectory—and if agentic workflows do become the default in software development, design, and operations—that assumption may prove prescient.
If not, Humwork has built an intriguing MCP server with a very specific 2 a.m. problem. Whether there are enough 2 a.m. problems to build a company around is the question the next twelve months will answer.
