# Moss

Aiolum story: https://aiolum.com/story/moss/ · added 2026-11-05 · updated 2026-11-06

**In short.** Moss (Real-time semantic search runtime for voice agents, United States): 803 GitHub stars and paid plans live, November 2026. Evidence grade E: no revenue figure was checked. Team: 1 named founder, founded 2024; 5 people on the Y Combinator profile, November 2026. Launched 30 October 2025; still running.

Real-time semantic search runtime for voice agents · United States · AI-native · team of 2 to 10 · still running · group: Developer tools and coding agents

- Outcome: 803 GitHub stars and paid plans live, November 2026 (evidence grade of the headline number: E; no revenue figure)
- Launch: 30 October 2025; months from launch to the outcome: 13
- Team: 1 named founder, founded 2024; 5 people on the Y Combinator profile, November 2026
- Outside funding: Y Combinator, Fall 2025 batch; amount not published
- Main factor: Latency
- Meters, 1 to 4: evidence strength 1, copyable pattern 1, capital needed 2
- Tags: AI-native, Small team, Funded, Running, B2B, Developer tools

Links:

- Visit moss.dev: https://www.moss.dev/

**The product.** A search engine library that a developer puts inside a voice agent, a copilot or a chat app. It turns documents into embeddings, keeps the index in the same process as the agent, and answers a query in under 10 milliseconds, without a call to a remote vector database.

**Why it worked.** A voice agent must answer within a fraction of a second, or the conversation feels broken. The founder writes that the slow step is almost always retrieval: each lookup crosses the network to a cloud database. Moss moves that lookup into the agent's own process.

Evidence grades: A = disclosed by a public company, a filing or a payment processor, or verified by one; B = two or more independent reports agree, or a journalist confirmed it; C = a founder claim that no one else verified; D = reported, not verified: one founder claim or a third-party estimate with no second source; E = no revenue figure found, so the headline is a non-money fact and the other facts carry their own grades; calc = computed on this page from graded facts.

## Summary

Moss sells semantic search, search by meaning instead of by exact words, as a runtime: a library that runs inside the customer's own program, in the browser, on a phone or on a server. Documents are turned into embeddings, lists of numbers that place texts with similar meaning close together, and the index of those numbers is loaded into memory next to the AI agent. The company says a lookup takes under 10 milliseconds, against hundreds of milliseconds for a cloud vector database, a database that stores embeddings and finds the nearest ones.

Sri Raghu Malireddi, a former machine-learning lead at Grammarly and Microsoft, founded Moss in 2024. It joined Y Combinator in the Fall 2025 batch and posted its launch on 30 October 2025, with 3 paying customers, 6 enterprise design partners and, in the founder's words, usage and revenue growing about 100% week over week. In November 2026 its Y Combinator profile lists 5 people, its main GitHub repository has 803 stars, and its site sells paid plans.

No revenue amount was found in any source, so this page has no money headline. The counts above are the company's own statements, except the GitHub numbers, which anyone can see.

## Pricing

Moss prices in November 2026, from the company's pricing page in Markdown (the HTML page loads its prices by script)

| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Developer (free) |  | 0 | 1 project, 3 indexes, 500 MB storage, 10 GB egress |
| Hobbyist |  | 30 | unlimited projects, 2 GB storage, 100 GB egress |
| Start-Up |  | 200 plus usage | 5 GB storage, 1 TB egress, bring your own embeddings |
| Enterprise |  | custom | SOC 2, HIPAA, data residency or VPC deployment |
| Moss Inference |  | usage only | 0.25 per 1M input tokens, 1.75 per 1M output tokens, 0.50 per 1,000 retrievals |

The pricing page shows monthly prices only and no yearly discount. Every plan includes unlimited local queries, because the search itself runs on the customer's machine; the plans differ in cloud storage, egress (data sent out of the cloud), ingest and control operations. Moss Inference, a hosted model called moss, is in private beta and is billed per token and per retrieval.

## How much revenue does Moss make? The numbers, graded

A five-person developer-tools company with paid plans and a public SDK repository, and no published revenue figure

| Fact | Value | Grade | Note | Sources |
|---|---|---|---|---|
| Founded | 2024, by Sri Raghu Malireddi (Founder and CEO) | C | Y Combinator company profile, read 5 November 2026. The launch post speaks of a founding team that knew each other for more than 8 years, but the profile names one founder. | [1, 2] |
| Team, November 2026 | 5 people | C | The team_size field of the Y Combinator profile, which the company fills in. The profile also lists 1 open engineering role. | [1] |
| Funding | Y Combinator, Fall 2025 batch | C | Y Combinator profile. No round amount was found in any downloaded source. | [1] |
| Public launch | 30 October 2025, Launch YC post with 107 votes | C | The created_at date and vote count of the launch post. | [2] |
| Customers, October 2025 | 3 paying customers, 6 enterprise design partners, 7 more evaluating | C | Founder's launch post. No customer is named as paying. | [2] |
| Growth, October 2025 | usage and revenue about 100% week over week | C | Founder's launch post. No base amount is given, so the rate cannot be turned into a figure. | [2] |
| Teams using it, November 2026 | 500+ teams, including Microsoft, EPAM, UC Berkeley, Carnegie Mellon, Podium and Stanford | C | Company documentation for developers. Not verified: no named organization confirms it on a page that was downloaded. | [5] |
| Main GitHub repository, November 2026 | 803 stars, 105 forks, 283 commits, BSD-2-Clause license | B | Visible on github.com/usemoss/moss. The repository holds the SDKs, a CLI, connectors and examples; the README tells users to point the SDK at Moss Cloud. | [6] |
| Query speed in the company's benchmark | 3.1 ms median against 351.8 ms for ChromaDB, 432.6 ms for Pinecone and 597.6 ms for Qdrant | C | Company benchmark on 100,000 documents, with a published script. The README says the other systems used an external embedding service and, for Pinecone and Qdrant, cloud search. | [3, 6] |
| Prices | free, then US$30 and US$200 a month plus usage | C | Company pricing page in Markdown, read 5 November 2026. | [4] |
| SDKs | Python, TypeScript and Node.js, a WebAssembly build for the browser, Elixir and C | B | GitHub README. The company says the core is written in Rust and WebAssembly. | [5, 6] |
| Time from launch to the November 2026 counts | 13 months | calc | Calculated: October 2025 to November 2026, counted by calendar month. | [2, 6] |

How to read this table: a grade C value is a statement by the founder or the company. A grade B value is visible on a public page that the company does not control (GitHub). A "calc" value was computed on this page. No revenue figure exists.

## When did Moss launch, and how did it grow? Timeline

From a 2024 start to a Y Combinator launch and a public SDK, in the company's own records. Each entry carries its grade.

- **2024** (grade C [1, 2]): Sri Raghu Malireddi founds Moss in San Francisco after work on large models and personalization at Microsoft and Grammarly.
- **Fall 2025** (grade C [1]): Moss joins Y Combinator's Fall 2025 batch.
- **30 October 2025** (grade C [2]): Launch YC post: 3 paying customers, 6 enterprise design partners, and usage and revenue growing about 100% a week, in the founder's words. SDKs for JavaScript and Python.
- **November 2026** (grade C [3, 4]): Paid plans from Hobbyist to Enterprise, and Moss Inference, a hosted model, in private beta.
- **November 2026** (grade B [1, 6]): The main GitHub repository has 803 stars and integrations for LangChain, LiveKit, Pipecat, Vapi, ElevenLabs and others. The Y Combinator profile lists 5 people.

## Why did Moss work?

### Latency

Moss sells speed in a place where speed decides whether the product works. In a voice pipeline, speech recognition, retrieval, the model and speech synthesis run one after another, and the company's FAQ says retrieval adds 100 to 500 milliseconds. Removing that network hop is the whole offer.

### Search inside the customer's process

The index is loaded into memory in the agent's own program, so a query needs no network. The cloud part only stores, syncs and distributes the indexes. That is why every plan has unlimited local queries and the price follows storage and data transfer instead of query count.

### Go where voice agents are built

The launch post names Pipecat and LiveKit, two frameworks for real-time voice agents, as close partners, and the repository ships ready examples for them and for Vapi and ElevenLabs. A developer who builds on those frameworks finds Moss as a plug-in.

## How it was built

These details come from the company's own pages and its public repository.

- The search core is built in Rust and compiled to WebAssembly, so the same index runs in a browser, on a device, at the edge or on a server.
- Built-in embedding models (the documentation names moss-minilm and moss-mediumlm), or the customer's own embeddings from any model, such as OpenAI's.
- Hybrid retrieval: semantic search and keyword search in one query, with metadata filters.
- SDKs for Python, TypeScript, Elixir and C, a command-line tool, an MCP server for AI coding tools, and connectors that ingest from SQLite, MongoDB, MySQL and Supabase.
- A managed cloud layer that stores indexes and syncs them to every running copy; indexes keep working offline.

Sources: [2, 5, 6]

## Who tried this before Moss, and why did they fail?

Three companies in our database of failed startups promised smarter retrieval or a voice assistant that runs locally, and failed. The values below are from that database and are not independently verified.

The failed startups come from our database of dead startups, retold here in our own words; Its entries are AI-assisted summaries of public sources, so they carry no evidence grade.

### Omnity (2015 to 2021; funding US$4.5M; cause: Competition)

- Promise: A semantic search platform that connects data across many fields.
- Why it died: Large companies with their own data and AI search took the market, and the cost of collecting and processing data was higher than the revenue.
- Record: 

### Petalite (2014 to 2025; funding US$12M; cause: Competition)

- Promise: A private voice assistant that runs on the device instead of in the cloud.
- Why it died: It sold hardware against Amazon, Google and Apple, which subsidize their devices.
- Record: 

### Answerly (2013 to 2016; funding US$550K; cause: Product/Tech Failure)

- Promise: FAQ answers matched to customer questions by language processing, inside company websites.
- Why it died: Language processing of the time could not handle real customer questions, and larger support tools already owned the market.
- Record: 

## What changed

Omnity and Answerly sold search by meaning before good embedding models were cheap, and Petalite tried to run a voice assistant on the device when that needed custom hardware. Moss sells neither a finished assistant nor hardware. It sells one component, retrieval, to developers who already build voice agents on large language models.

Two things are new. Small embedding models now run in a browser or on a phone, and voice agents created a buyer for whom 300 milliseconds is a product defect. Moss's risk is the one that ended Omnity: larger database and cloud companies can add a local mode of their own.

## Can one person copy Moss now?

**Verdict: Hard for one person.**

The product is low-level systems work, a Rust and WebAssembly search engine with SDKs in several languages, built by a founder with years of machine-learning work at large companies and a team of five. What a small builder can copy is the choice of problem: find the one slow step in a new kind of AI product and sell only that step.

1. Measure where time goes in a pipeline that many teams are building now (here, voice agents).
2. Build the fix as a library that runs in the customer's own process, so there is no new service for them to operate.
3. Ship ready examples for the frameworks those teams already use.
4. Publish a benchmark with its script, so buyers can repeat it.

**What is still open.** Other steps on the critical path of real-time AI: speech recognition on the device, caching of model answers, and turn detection in voice calls.

## What could go wrong for Moss?

- **What is not verified.** No revenue figure was found. The only money-related statements are the founder's in the 30 October 2025 launch post: 3 paying customers and usage and revenue growing about 100% week over week, with no base amount. No payment-processor profile, filing or interview with a figure was found. A reader could look for a revenue figure in a later founder interview or a funding announcement.
- **Customer claims and logos.** The developer documentation says Moss is trusted by 500+ teams, including Microsoft, EPAM, UC Berkeley, Carnegie Mellon, Podium and Stanford, and the home page shows a row of production users. No downloaded page from those organizations confirms it. A counter of conversational minutes served on the home page loads by script, so its value could not be read. The quotation on the home page is from the CEO of Y Combinator, which invested in Moss.
- **Competition.** On this site, Chatbase and DocsBot AI sell finished chatbots built on retrieval; Moss sells the retrieval layer underneath such products. Its own pages compare it with Pinecone, Qdrant and ChromaDB, and its benchmark is its own: the README says the other systems were measured with an external embedding service and cloud search.
- **Dependence on model providers and voice platforms.** Moss ships its own embedding models and lets customers bring embeddings from OpenAI or others, so search itself does not need a model provider. Its buyers do: the voice agents it serves run on outside language and speech models, and its distribution goes through frameworks such as LiveKit, Pipecat, Vapi and ElevenLabs, which could add their own retrieval.
- **Team and funding.** Moss is a Y Combinator company with venture funding and a hiring page, not a bootstrapped one. The Y Combinator profile lists 5 people in November 2026; the size of any round was not found.

## Sources

1. Y Combinator company profile: Moss, Real-time semantic search for Conversational AI — https://www.ycombinator.com/companies/moss (grade C). Read in full, including the page data with year_founded, team_size, batch and founder. Downloaded by the pipeline on 5 November 2026. Company-entered fields.
2. Launch YC, 30 October 2025: Moss, Real-time Semantic Search for Conversational AI — https://www.ycombinator.com/launches/Oiq-moss-real-time-semantic-search-for-conversational-ai (grade C). Read in full. Founder's launch post. Source for the paying customers, design partners, growth rate, partners and the founder's background.
3. Moss home page — https://www.moss.dev/ (grade C). Read in full on 5 November 2026. Source for the latency claims, benchmark chart, integrations and FAQ. Some counters load by script and could not be read.
4. Moss pricing page in Markdown — https://www.moss.dev/pricing.md (grade C). Read in full on 5 November 2026. The HTML pricing page loads its prices by script; the Markdown version linked from llms.txt gives them.
5. Moss developer documentation for AI coding agents (llms-full.txt) — https://moss.dev/llms-full.txt (grade C). Read the overview, installation and bring-your-own-embeddings sections. Source for the 500+ teams claim and the embedding model names.
6. GitHub repository usemoss/moss — https://github.com/usemoss/moss (grade B). Read the repository page and README on 5 November 2026. Source for stars, forks, commits, license, SDKs, integrations and the benchmark table.
7. Database of failed startups, local export — /method/#dead (grade C). Local export. AI-written summaries, not independently verified. Source for the failed predecessors.

Contact: aiolum@livyden.com. Run Moss? Send corrections or a verified figure to aiolum@livyden.com (subject: Correction: Moss).