Mercor matches hundreds of thousands of domain experts with frontier AI labs that need human data for training and evaluation, and also runs large batches of automated grading jobs for labs and enterprises. Its load is bursty: model traffic can sit at peak capacity for days while individual evaluation or labeling jobs are due within 24 hours, so requests stalled behind a provider's rate limits put client deadlines at risk. Business teams, meanwhile, need per-engagement applications that let experts see their tasks and progress without waiting on changes to the core platform. Mercor addresses the first problem with Vercel AI Gateway and the second with Vercel Passport.

Token throughput as a delivery constraint

Mercor's earlier combination of direct integrations and third-party gateways lacked the token-per-minute capacity its open-model workloads required during surges. Requests that hit provider limits landed in retry queues, and batch jobs that should have closed inside a client's 24-hour window ran past 12 hours — more than half the deadline consumed.

AI Gateway added inference capacity and failover routing across providers when rate limits were reached, so batch work could keep moving instead of waiting for retries. Mercor reports those jobs now complete in 2 to 3 hours. In the first 10 days on Gateway the team routed 18 million requests and 63 billion tokens without hitting rate limits. Mercor still runs multiple gateways and vendors, and values Vercel for the combination of model access, controls and support when additional capacity is needed. Engineers also gain a single API spanning major foundation model providers and a growing open-model catalog, so a new model can be tested without a new provider integration.

Spend and usage controls for hundreds of engineers

Growth in model use exposed a visibility gap: the previous setup offered little insight into token spend by team or project, and one erroneous loop in a script could drain the available budget. Hundreds of Mercor engineers now reach open models through AI Gateway using per-user keys with expiration dates, token limits and spend limits. Security can grant experimentation a budget and an end date, then use usage records to identify which user, project and workload drives consumption. Engineers can switch models behind the same gateway under those controls rather than managing separate provider credentials.

Applications built outside engineering

A few hundred people across Mercor's business teams now build on Vercel, and the company's project count has gone from fewer than a dozen to roughly 2,000 — many of them temporary environments tied to a single client engagement. One production case gives experts detailed views of their project tasks and progress; instead of asking engineering to add each engagement's interface to the core platform, business teams build the interfaces and share them with the experts doing the work.

Sharing requires control over sign-in without standing up separate authentication for every project. Passport supplies a common authentication layer that handles sign-in to Vercel deployments through an existing identity provider. Mercor's identity team built a product called Audiences that integrates Passport as a first-class primitive to automate provisioning: once a team selects the experts who need access, Audiences creates the IdP group, the Vercel connection and the Passport configuration. Mercor reports provisioning in seconds for audiences from dozens of experts to tens of thousands. Offboarding runs the same workflow in reverse — because access derives from the IdP group, deleting the group at the end of an engagement removes project access, leaving no stale permissions.

Scoped permissions for internal agents

Mercor also runs agents that answer questions from marketplace experts, automate HR workflows and help employees navigate internal systems. Each agent gets narrow permissions matched to its job rather than inheriting everything available to the person who invoked it. For agents on open models through AI Gateway, Mercor scopes model access, sets expiration dates and caps spending, separately from the systems and actions each agent may reach.

Where the platform is heading

Mercor expects more open models behind its internal agents and systems, and more non-engineering teams building tools for experts. The goal is to keep teams from spending their time on model integrations and application authentication so they can answer client workloads with tight deadlines — making the secure path the fastest one, and supporting growth without security work rising at the same rate.