Minicor Raises $3.8M Seed for AI Agents [2026]

Minicor, a San Francisco startup co-founded by Nigerian-born software engineer Saheed Akinbile, announced a $3.8 million seed round on October 14, 2026 The company builds technology that lets AI applications work with older business software, especially desktop systems that never shipped with application programming interfaces (APIs). Brad Gillespie of IA Ventures led the round

The story spread fast in Nigerian tech circles because of the founder’s path. Akinbile and co-founder Faizaan Chishtie got into Y Combinator’s Spring 2026 batch after seven applications and three interview invitations. Some headlines frame that history as seven rejections. The available reporting confirms the seven applications, but it does not say each one ended in a rejection, so this article sticks to what the record supports.

Beyond the founder story, the deal matters for a plain commercial reason. Frontier models keep getting better at reading screens and using tools, as our coverage of GPT-6.1 Terra showed. Yet a model cannot help a hospital, a freight broker or a manufacturer if the software holding their data has no door for it to knock on. Minicor is betting that building the door is a bigger business than building another model.

Don’t miss new tech stories on Google

Add Tech Insider once in the Google app and our stories appear in your news suggestions.

What Minicor Announced on October 14

The announcement centers on one product idea. Minicor sells tooling that lets AI agents interact with legacy desktop software without replacing it. According to the reporting, developers can convert desktop workflows into API endpoints, so newer applications can talk to existing systems through a standard interface.

That framing is deliberate. Most enterprise buyers do not want a rip-and-replace project. They want the old billing terminal, the warehouse client or the clinical records app to keep running while new AI features sit on top. A workflow turned into an endpoint gives developers something familiar to call, with inputs and outputs they can test, instead of a fragile chain of simulated mouse clicks.

The funding itself is modest by 2026 standards. A $3.8 million seed would not cover a single large GPU cluster for long. But the product does not need one. Minicor sells software that sits between existing models and existing applications, and that is a capital-light position compared with training frontier systems.

Who Is Behind Minicor

Saheed Akinbile is the Nigerian-born engineer at the center of the coverage, and Faizaan Chishtie is the other co-founder named in the reports. The company is based in San Francisco. Beyond those names and locations, the available reporting offers few biographical details, and no complete on-record quote from either founder appeared in the

That gap is worth stating plainly. Several summaries of the deal paraphrase the founders’ views, but none of the verified material gives their exact words. We therefore describe the product from the announcement and from third-party funding trackers, and we avoid putting words in anyone’s mouth.

The Nigerian angle gives the story reach on the continent. African outlets have long tracked founders with roots in Lagos, Abuja and other hubs who build companies abroad, and the Minicor round fits that pattern. The company is not a Nigeria-based startup, though. It operates from San Francisco, and its customers are presumably enterprises that run old desktop software wherever they sit.

The Seed Funding Round and Its Investors

Brad Gillespie of IA Ventures led the $3.8 million round. The list of other participants is longer than the check size suggests: Y Combinator, Paul Graham, Pioneer Fund, Robinhood Ventures, Garage Capital, Rebel Fund, K5 Global and 20VC. Funding tracker Signalbase lists the deal under the headline “Minicor Secures $3.8M,” and StartupDB keeps a page of Minicor’s backers.

Y Combinator and Paul Graham both appear on the list, so the accelerator that admitted Minicor in the Spring 2026 batch also put money into the round. The reports say nothing about how much each investor contributed, so the lineup tells us who is involved but not who holds the most weight.

The reports do not disclose a valuation, ownership split or runway. We will not guess at them. Readers should treat any number floating around social media for those items as unconfirmed until Minicor or its investors publish them.

The Y Combinator Path: Seven Applications, Three Interviews

The detail that drove the headlines is the Y Combinator history. Minicor’s admission into the Spring 2026 batch followed seven applications and three interview invitations. For founders, that sequence is the real lesson. Getting three interviews from seven tries means the application kept improving, or the product did, or both.

The “seven rejections” framing in some coverage is catchy but unverified. The report confirms the number of applications. It does not state that every one was turned down, and an application can end in other ways, such as being withdrawn or superseded by a newer submission. Careful readers should hold the count at seven applications and three interview invitations.

Persistence stories sell well because they flatter a common experience. Plenty of technical founders apply more than once to top accelerators, and the model has long rewarded teams that return with sharper products. Still, an accelerator slot proves nothing about product-market fit. The more telling evidence will be paying customers, which the announcement does not quantify.

Why Legacy Desktop Software Still Blocks AI Agents

Enterprise IT carries a long tail of programs built before web APIs became standard. Many run on Windows desktops, store data in local databases and expose nothing except a graphical interface. Healthcare records tools, accounting packages, logistics clients and document managers often fit that description. An AI agent that can browse the web and call cloud services hits a wall when the task requires one of these programs.

Replacing them is rarely quick. Migration projects stretch for years, cost real money and carry operational risk, so companies tolerate old systems far longer than engineers would like. Meanwhile, executives want AI features now. That mismatch creates the opening Minicor targets.

The agent-security angle matters here too. Giving an autonomous system access to a desktop that holds sensitive records raises hard questions about identity and permissions. We looked at that problem in our comparison of the identity providers for AI agents, and the same concerns apply to any tool that logs into old software on an agent’s behalf.

How Minicor Turns Desktop Workflows Into API Endpoints

The reporting describes the core mechanism in a single line: a desktop workflow becomes an API endpoint. Developers define or record a task inside the old application, and Minicor exposes that task as a callable interface. A newer application, or an AI agent, then calls the endpoint and receives a result, without knowing a Windows program did the work behind the curtain.

From screen clicks to callable endpoints

Think of an insurance clerk who opens a claims program, searches a policy number, copies three fields and closes the window. Minicor’s approach captures that sequence once. After that, a developer calls an endpoint with a policy number and gets the three fields back. The same pattern could apply to invoice lookups, inventory checks or patient record queries.

The practical gain is testability. An endpoint has defined inputs, defined outputs and a clear failure state. A free-roaming agent that improvises each click does not offer those guarantees. Engineers who have debugged flaky UI automation will see why that difference matters.

Where recent model advances fit

New models sharpen the case for this kind of plumbing. Reports describe OpenAI’s GPT-6.1 Terra, released in October 2026, as focused on visual reasoning and computer use, and NVIDIA’s Nemotron 3.5 Super VL was reported out on October 15. Anthropic’s current lineup, covered in our Claude Opus vs Sonnet vs Haiku 5.5 breakdown, also pushes agentic work. The availablet the link as market context rather than a product claim

Minicor vs RPA vs Computer-Use Agents

Minicor does not enter an empty field. Robotic process automation vendors have automated desktop tasks for years, and model labs now ship agents that operate a computer directly. The table below is our own qualitative framing of the approaches, not a benchmark, and it uses no measured performance numbers.

The interesting question is whether Minicor sits closer to RPA or to agents. By the company’s account, its agents learn how the target software works, build the automations and keep them running. That blends both worlds. The model-driven part handles setup and repair, while the output behaves like a conventional integration that developers can call on demand.

Incumbent RPA vendors will not sit still. They hold large customer bases and established procurement channels, and they can bolt language models onto existing tools. Model labs, for their part, can improve computer-use accuracy until a middle layer looks unnecessary. Minicor has to prove the middle layer pays for itself.

What Is Confirmed and What Is Not

Funding stories attract inflated numbers, and this one is no exception. A product-usage figure of 2.5 million automation executions has circulated, and one it below as unconfirmed and keep it out of our analysis

The distinction between application count and rejection count may look pedantic. It is not. A fabricated hardship detail can spread faster than the correction, and trade press that repeats it without checking erodes trust for founders who tell their stories accurately.

Market Impact: Where Agent Infrastructure Money Is Flowing

Investors in 2026 keep funding the layers around the models. Security for agents, identity for agents and now access for agents all draw checks. Our report on Infisical’s $40 million Series B for AI agent security showed one slice of that trend. Minicor’s seed targets another slice: the unglamorous work of connecting agents to the software companies actually run.

The pattern also shows up in the weekly deal flow. Our roundup of the biggest funding rounds in physical AI and beyond tracked how capital is spreading from model builders toward the companies that make models useful. At $3.8 million, Minicor sits at the small end of that spectrum, but small infrastructure rounds often precede larger ones when early customers appear.

For enterprise buyers, the impact is practical. If tools like Minicor work as described, a company can pilot AI agents against its oldest systems without waiting for an ERP migration. That could shorten AI project timelines from years to months in some cases. We stress “could,” because no customer results have been published.

Historical Context: From Screen Scraping to Agent Tooling

Connecting new software to old software through the screen has a long history. Terminal emulators and screen scrapers fed mainframe data into newer systems decades ago. RPA turned the idea into a product category by letting business analysts record tasks and replay them. Each wave promised to spare companies from rewriting core systems, and each wave hit the same wall: automations broke when a screen changed.

Language models changed the economics of repair. A model can look at a changed screen, work out what moved and adjust the automation. Minicor describes its automations as self-healing for that reason, in line with its product materials. The claim is plausible in principle, and it is also the hardest one to verify from outside. Self-healing only counts when it works at 3 a.m. on a quirky application nobody documented.

The other shift is how developers consume the result. Instead of a bot that runs on a schedule, they want a tool an agent can call. That mirrors the broader move toward standard tool interfaces for agents, where a service describes what it can do and any compliant agent can use it. Turning a desktop app into such a tool is the logical next step.

Risks and Open Questions for Enterprise Buyers

Security and access control

Any layer that logs into a desktop application on someone’s behalf inherits that application’s credentials and data. Buyers will ask where credentials live, who can trigger an endpoint, and how actions get logged. Regulated sectors such as healthcare and finance will demand audit trails before they allow an agent near patient or account records. The announcement does not detail these controls, so buyers should request them in writing.

Code produced by AI tools also needs scrutiny. Our look at AI-generated code security tests found that top models still fail a meaningful share of checks. If automations are generated by agents, security teams should review them like any other code.

Reliability and vendor risk

Desktop automation fails in odd ways, from pop-up dialogs to license prompts to a login screen that appears once a month. A seed-stage startup has had little time to meet those cases at scale. Companies that depend on a single small vendor for a critical workflow should plan fallbacks and keep manual procedures alive during any pilot.

There is also platform risk. If major model providers fold desktop control into their core products, a standalone layer faces pricing pressure. Minicor’s defense would be depth: tooling, monitoring and repair that general-purpose agents do not ship by default.

Five Predictions for Agent-to-Desktop Tooling

These are our analysis, not reported facts, and none comes from Minicor.

• More seed rounds will target legacy access. The pain is widespread and the capital need is low, so expect additional startups to pitch the same bridge between agents and old software through 2027.
• Incumbent RPA vendors will rebrand around agents. Expect product launches that pair existing bot libraries with language models and call the result agent-ready.
• Buyers will demand audit logs first. Security reviews, not demos, will decide which tools reach production in regulated industries.
• Model labs will narrow the gap. Better computer-use accuracy will reduce the need for hand-built fixes on simple tasks while complex, high-volume workflows keep favoring dedicated tooling.
• Customer proof will matter more than funding. The next signal to watch is named enterprise deployments, not another investor on the cap table.

For ongoing coverage of model releases, funding and agent tooling, follow our AI and machine learning section.

Frequently Asked Questions

What is Minicor?

Minicor is a San Francisco startup building technology that lets AI applications interact with older business software, particularly desktop systems without APIs. Developers can convert desktop workflows into API endpoints so newer applications can communicate with existing systems.

How much did Minicor raise and who led the round?

Minicor raised $3.8 million in seed funding, announced on October 14, 2026. Brad Gillespie of IA Ventures led the round, with participation from Y Combinator, Paul Graham, Pioneer Fund, Robinhood Ventures, Garage Capital, Rebel Fund, K5 Global and 20VC.

Who founded Minicor?

Coverage names Saheed Akinbile, a Nigerian-born software engineer, and Faizaan Chishtie as co-founders.

Did Minicor’s founders really get rejected seven times by Y Combinator?

Not confirmed. The report says Minicor entered Y Combinator’s Spring 2026 batch after seven applications and three interview invitations. It does not clearly state that all seven applications were rejections.

Why can’t AI agents already use old desktop software?

Agents work best through APIs, which give them structured ways to read and write data. Many legacy desktop programs expose only a graphical interface, so an agent must either simulate a person at the screen or rely on a layer such as Minicor’s that wraps the workflow in an endpoint.

Has Minicor published customer numbers or usage figures?

The verifiedon automation executions has circulated, but onet

Is Minicor a Nigerian company?

No. The company is based in San Francisco. Its co-founder Saheed Akinbile is Nigerian-born, which is why Nigerian outlets such as The Nation and PM News Nigeria covered the round.

Does Minicor use GPT-6.1 Terra or another specific model?

The availableant context for the category, but the reporting does not link Minicor to any particular one