tl;dr: the official Y Combinator startup directory lists 581 companies across its live batches, and 6,124 in the full back catalog. the ten companies below are in none of them. they are in frontrun, some since June. one of them has four followers.
the YC startup directory is published, not live
there is a gap between being in a Y Combinator batch and being in the Y Combinator directory.
a founder gets in, moves to San Francisco, starts building and starts hiring. the directory entry arrives later, on YC's publishing schedule. in between there is a real company with no public record: no directory listing, no press, often no launch. that window is the only time an investor can reach a YC founder before every other investor has the same list.
this post is what that window looked like for YC S26 on august 7 2026.
how frontrun tracks new Y Combinator entrants
frontrun started as one signal: watch the follow graphs of 2,000+ venture investors on X and flag the accounts they quietly converge on. that is still the core, but it is no longer the only input.
the pipeline now also runs accelerator-batch monitoring, launch and fundraise detection, LinkedIn stealth-profile signals, and a founder resolution layer that ties a person to the company they are actually building. a company can enter from any of them, and whichever fires first stamps the date.
that matters for a batch like this one, because a stealth YC founder often has no company account at all. five of the eleven below still do not. what they have is a founder, a domain, and a few investors who found them early.
the same pipeline runs against every incoming batch. YC Fall 2026 and YC Winter 2027 entrants are being flagged now, with dates, months before the YC startup directory publishes them.
the list
1. Tokenless: a model router that halves your inference bill
Tokenless (@usetokenless) · 21 followers · flagged july 16
a drop-in replacement for your OpenAI and Anthropic API calls. it fans each request out to a group of models, watches them think, selects the one that is clearly on track and cancels the rest, so you only pay for what you needed.
founders: @rohit_writes (CS PhD candidate at Princeton PLI, prev UC Berkeley EECS) and @landrew7363 (27 followers)
2. Super Rad Company: secured agentic computing
Super Rad Company (@superradcompany) · 74 followers · flagged july 13
their manifesto: agents are becoming the interface to computing, and the infrastructure to run them in isolation with fine-grained policies is still in its infancy. so they build for federation, isolation and trust, in public and open source. two things shipped: microsandbox, hardware-level isolation for untrusted agent-generated code, and tool.store, a package manager for MCP tools.
founders: @toksdotdev (cofounder of microsandbox, prev Cloudflare, 125 followers) and @asciidotdev
3. Princeps: compute finance
Princeps (@princepsdev) · 84 followers · flagged july 13
the site is a wall of scrambling characters that resolves into three things: a ComputeBond pricing class, a CUDA settle kernel that multiplies hours by price, and an sla_breach function that turns downtime below a floor into a penalty. that is the whole thesis, written in code. the bio reads "manifest density. sovereign compute."
founder: @anya_trofimova7 (VP of the Oxford Union, prev Palantir and Chatham House)
4. Hebbian Robotics: data curation for robotics
Hebbian Robotics (@hbr_pbc) · 205 followers · flagged june 23
"robotics teams are collecting more data than ever, yet much of it never reaches training." curation is manual and quality is rarely defined well enough to be reproducible, so teams throw away most of what they collect. their answer is Pareto, high-quality data curation for robotics, treating datasets with the same rigor researchers give models. the site says backed by Y Combinator plus angels from Oracle and Google DeepMind.
founders: @bdono_ (robotics PhD at Columbia, NTU Singapore) and @kstonekuan (prev Jane Street, Verkada)
5. Blaise: enterprise software, built by an AI system
Blaise (@blaisedotnew) · 395 followers · flagged july 1
Blaise builds all your enterprise software, deployed on your own infrastructure, "doing the work of an entire consultancy, but faster." Stockholm based. "your ideas, to production. by swedish superintelligence."
founder: @alexwikstrom (CEO, ex Lovable, prev physics)
6. Inner: supply chain security for AI coding agents
Inner (@securewithinner) · 4 followers · flagged july 23
it does not stop at reading code, it runs it: every npm and PyPI package executed inside an AI sandbox before your coding agent installs it. one command puts Inner between the agent and the registry. it routes npm, pnpm, yarn, pip, poetry and uv, and wires itself into Claude Code and Cursor over MCP. built by a team of Harvard and MIT engineers.
founder: @IKhaliliya
7. CueBench: RL environments for what frontier models fail at
CueBench · no company account yet · flagged august 3, at 9 followers
"we think the next step toward AGI will come from training models in highly specialized environments, especially the ones where frontier models still fail more than 80% of the time." pre-training is delivering diminishing returns, so they build environments for the things models are worst at: scientific reasoning, experiment design, and inference engineering.
founder: @neelapolitan
8. Kandor: the revenue cycle for specialty clinics
Kandor · no company account yet · flagged july 8, at 78 followers
"Kandor runs the operations behind specialty clinics: the phones, the texting, the eligibility checks and the whole billing cycle." two products, Kandor Suite for the team and Kandor Billing as a service. the pitch underneath it: coding and denials eat up to a fifth of a practice's revenue and most denied claims never get touched, so Kandor catches undercoding, fights denials, and recovers abandoned revenue autonomously. "focus on the patients."
founders: @devansh_1704 and @cyrusnaficy
9. Veltha: AI that knows the insurance regs
Veltha · no company account yet · flagged july 8
"AI that knows the regs better than your best adjuster." every claim checked against the actual regulation, with paragraph-level citations, before errors become audit findings. two lines to start: crop insurance checked against FCIC, LAM and PPSH, and workers' comp compensability, with an audit trail under every decision.
founder: @edin_citaku
10. Shepherd: one memory for every technical domain
Shepherd · no company account yet · flagged july 11, at 168 followers
research, hardware, firmware, software and ops in a single shared memory your team can trace and your agents can act on. every design review, experiment, message and commit goes in. integrations run from Slack, Notion and Linear to raw schematics.
founders: @stochasticsloth (CTO, Stanford) and @ishanr07 (CPO, Stanford)
11. E.env: markets as RL environments
E.env · no company account yet · flagged august 1
market-derived, non-saturating environments that teach agents applied ML and long-horizon planning under adversarial noise. agents get professional tools plus bash to build their own, and have to actually trade. the thesis: "static synthetic benchmarks are unrealistic and saturate quickly, whereas markets are non-saturating and self-improving."
founder: @rricefan
four followers is the whole point
Inner has four followers. CueBench was flagged at 9. Kandor at 78.
by any audience metric those numbers are nothing, which is exactly why follower count is the wrong thing to watch. a stealth founder has no audience by construction. what they have is a small number of people who already know, and those people leave a trace before the founder does.
tracking YC Fall 2026 and Winter 2027
if you found this page looking for a YC S26 company list, the more useful list is the one that does not exist yet.
this was not a one-off scrape. the pipeline runs every day against every incoming Y Combinator batch, so Fall 2026 and Winter 2027 founders are being flagged right now, at double-digit follower counts, months before the Y Combinator startup directory publishes them.
everything above came out of the frontrun API and MCP server.
the directory will catch up. the graph was first.
see who is on the list today →
read next: a16z speedrun cohort 007, all 29 companies · startups flagged before they raised · how vcs find startups before they raise
open source deal flow
companies and founders surfaced on frontrun. curated, every week.
