Clipto’s $250 million bet on local AI search says the file pileup is the real problem

Generative AI has made it easy to create more stuff than anyone can keep track of. Clipto is betting the pain point is retrieval.

The San Francisco startup, founded in 2023, just raised $15 million in an all-equity round at a $250 million post-money valuation. The investor list includes HSG, GL Ventures, EnvisionX Capital, Palm Drive Capital, Hans Tung, Lu Zhang, and 522 Ventures. Clipto says it’s already at $15 million in annual recurring revenue and profitable on a net-income basis. For a two-year-old company with just over 20 employees, that’s a strong result.

The product is built for a problem that gets worse as every laptop, desktop, and external drive turns into a junk drawer. Clipto indexes videos, audio, images, meetings, and documents stored on a user’s computer. Instead of digging through folders or guessing filenames, users can search in plain language or ask ChatGPT, Claude, and other AI tools to find files for them.

That sounds simple. It isn’t.

Searching your own files is harder than it looks

Consumer file search has been around forever, and it’s usually bad. Traditional indexing can tell you where a file lives and maybe what words are inside it. That falls apart fast for media. A video clip doesn’t have useful text unless someone transcribed or tagged it. An image needs OCR, object detection, facial or scene recognition, maybe all three. Audio needs speech-to-text plus speaker segmentation if you want anything better than a wall of timestamps.

Clipto’s pitch is that modern AI finally makes this workable at scale on personal hardware. That matters. A lot of AI search products quietly depend on cloud processing, which is fine until users ask where the data goes, who can read it, and what happens when sensitive documents leave the device.

Clipto says all processing runs locally on the user’s machine without cloud services. That’s the part to focus on. Local-first AI search gives the company a cleaner security story and cuts recurring inference costs. It also makes the engineering harder. The models have to be small enough to run on consumer machines, fast enough to stay usable, and accurate enough to avoid feeling flimsy.

That trade-off is ugly. If you want real-time indexing across terabytes of mixed media, you need a pipeline that can handle embeddings, transcription, metadata extraction, and retrieval without turning a laptop into a heater.

Platform vendors already crowd this space

Clipto isn’t walking into a blank market. Adobe, Apple, and Google all have versions of AI-powered search inside products they already own. Adobe has media intelligence in Premiere. Apple Photos can find images and videos with natural-language queries. Google Photos does something similar. Those companies already sit close to the user’s content, which is the main advantage here.

So a standalone product has to justify itself. Clipto says the difference is scope. It can search across videos, audio, images, and documents, while the platform vendors mostly stay inside their own ecosystems. It also lets users expose selected information to external AI tools like ChatGPT and Claude.

That last part is probably the strongest thing it does, and the riskiest.

A lot of people want a private content index that can still feed an assistant on demand. The obvious problem is access control. If an AI agent can query your whole archive, it can surface the wrong thing at the wrong time. Clipto says access to indexed files requires the user’s active request and authorization, and that any AI application using its indexed data can only retrieve information within the scope the user specifies.

That’s the right direction, but the security bar is high. Scoped access sounds clean until you try to define it across a messy archive of meeting notes, drafts, clips, scans, screenshots, and legal docs. Fine-grained permissions, auditability, and revocation matter a lot more than demo polish.

MCP is the practical part

Clipto added support for Model Context Protocol, or MCP, about two weeks ago. For developers, that’s the detail worth watching.

MCP has become a standard way to connect AI apps and agents to outside tools and data sources. In practice, it gives a model a structured path to ask for data it’s allowed to use, instead of relying on one-off plugin integrations or custom wrappers for every app. It fits the current agentic AI wave, where systems need controlled access to user-owned context.

For Clipto, MCP is as much a distribution move as a technical one. It makes the product easier to plug into the AI tools people already use. It also lowers the odds that Clipto becomes a dead-end app with a nice interface and no ecosystem. If Claude, ChatGPT, or another agent can query the local index through a standard protocol, Clipto becomes part of the workflow instead of another tab people forget to open.

It also raises the stakes. Once an index can be queried by outside agents, authorization, logging, and context scoping stop being nice-to-haves. They’re the product.

Henry Kang has been circling this problem for years

Clipto founder Henry Kang didn’t stumble into this category by accident. He says he’s been working on related ideas for roughly two decades. Back in 2006, as a PhD student at Carnegie Mellon, he worked on robots that could record their surroundings, identify objects, and remember where to find them. Later he built a startup for closet inventory and outfit suggestions. His second company, Zenvideo, focused on video creation and was acquired by Tencent in 2020.

That background explains why Clipto is centered on retrieval instead of generation. Kang’s basic argument is that the AI era hasn’t created a content shortage. It’s created a content glut. Video footage, recordings, screenshots, and documents are piling up faster than people can label or reuse them.

That’s a believable thesis. It fits how knowledge work actually happens now. Teams don’t lack output. They lack memory. Meetings get recorded but never revisited. Drafts get saved in three different places. Research notes disappear into folders nobody wants to open. The AI layer people want most often isn’t a chatbot. It’s a decent index.

The customer mix is wider than creators

Clipto started with video creators managing footage across computers and external drives. That makes sense. Creators already feel the pain of asset sprawl, and video is one of the hardest file types to search well.

But Kang says creators now make up only about a quarter to a third of users. The rest are lawyers, doctors, researchers, marketers, HR professionals, professors, and students.

That broader mix matters. It suggests the product is moving from media management into general knowledge retrieval. The use cases are obvious:

• lawyers looking for the clip where a witness said something specific
• doctors searching recorded consultations or scans
• researchers digging through lectures and interviews
• marketers finding product footage or campaign assets
• HR teams retrieving policy files and training videos

Each of those groups has different tolerance for latency, accuracy, and privacy. A creator can live with an imperfect result if it saves time. A legal or medical user is less forgiving. That puts a ceiling on the product. If Clipto wants to keep moving into professional workflows, it needs more than decent semantic search.

The economics look good, but hardware is still a problem

Clipto says the new funding will go toward AI models and computing infrastructure needed to run on consumer hardware, plus more integrations with AI agents. That wording is doing a lot of work.

Consumer-device inference sounds efficient, and in some ways it is. It avoids cloud bills and keeps data local. But it also means the company has to support a messy range of CPUs, GPUs, memory sizes, and operating systems. Running multimodal search on a MacBook Pro is one thing. Running it on a thin laptop with limited RAM is another.

That’s where a lot of local AI products get messy. They either tune hard for high-end machines, which narrows the market, or they cut model quality and end up with search results that feel vague and unreliable.

There’s also the scale issue. Indexing terabytes of mixed media isn’t free even when inference happens locally. The system still needs to chunk, embed, transcribe, and update metadata efficiently. If indexing is slow or battery-hungry, users will notice. They usually do.

The valuation says investors want sticky AI utilities

A $250 million valuation for a company at $15 million ARR is rich, but not absurd in this market if retention is strong and the product is sticky. Search tools can become deeply embedded. Once someone’s archive is indexed and tied into their workflow, ripping it out is annoying enough that many people don’t bother.

That stickiness is the point. If Clipto can become the memory layer for personal and team files, it has a real shot. If it stays a smart way to search old videos, it gets boxed in by platform vendors with bigger distribution and bundled pricing.

The bigger question is whether AI file search becomes its own category or just a feature. Right now, the answer looks mixed. Adobe, Apple, and Google are already building it into their ecosystems. That’s the pressure. Clipto is betting people will still pay for a product that works across everything, stays local, and plugs into the AI tools they already use.

That’s a reasonable bet. It’s also the kind that gets harder every quarter as the giants copy what works.