Moss growth playbook

Render the four plans (Developer free, Hobbyist $30, Start-Up $200 plus usage, Enterprise custom) and the Moss Inference token prices on the server, in the first HTML response of /pricing. Keep the interactive parts as an enhancement. Agree one set of plan facts: credits or limits, concurrent sessions, SLA and support. Generate the pricing page, pricing.md and the pricing section of the agent reference from that single file, so they cannot drift. Add a plain-text FAQ (what "unlimited local queries" means, when overages apply, how to move up a plan), a meta description that names the free plan, and Offer markup for each plan. Then check what a crawler, a link preview and an AI assistant see.

Developers compare retrieval layers on price before they sign up, and a growing share of that research goes through crawlers and AI assistants that read HTML or Markdown, not rendered JavaScript. A visible free plan answers the first objection, and one consistent set of facts stops assistants quoting limits that Moss does not offer. The fix is a rendering change plus one shared data file, so it is cheap for a five-person team.

[ ] Plan names and prices in the first HTML response of /pricing
[ ] Free plan named in the title or meta description
[ ] One data file feeds /pricing, pricing.md and the agent reference
[ ] Credits vs limits, concurrent sessions, SLA and support agreed for every plan
[ ] Moss Inference token prices shown
[ ] Plain-text FAQ: local queries, overages, upgrades
[ ] Offer markup for each plan
[ ] Checked: no-JavaScript fetch, Slack/X preview, two AI assistants

• Published limits commit Moss to numbers it may want to change

  Date the page and change the single source file, so every version updates together.

• Server rendering slows the page

  The plan cards are static text; render them statically at build time.

• Your websitePricing page, fetched twice on November 9, 2026: the page title is "Pricing | Moss", but the page text has no plan names, no dollar amounts and no mention of the free plan or credits. Below the "Pricing" heading there are only logos, the SOC 2 Type 2 and HIPAA badges and the footer, so the plans appear to load only in the browser.Source
• Site profileAn automated profile of moss.dev (November 10, 2026) read the pricing page as having no starting price, no tier names and no free tier, although the main button on every page is "Start Free".Source
• Your websiteThe Markdown version of the homepage that moss.dev returns to text readers ("Moss — Complete Technical Reference for AI Agents") describes the plans differently: Developer "Free" with "$5/month in credits"; Hobbyist $30 a month with "3 concurrent sessions" and file uploads; Start-up $200 a month with "100 concurrent sessions" and email support; Enterprise with a "99.9% SLA", SSO and "24/7 Slack support". It also says "Trusted by 500+ teams" and names InferEdge Inc., San Francisco.Source
• Your websiteMarkdown pricing file linked from llms.txt: Developer (Free), "1 project, 3 indexes per project", "500 MB storage, 10 GB egress", "Unlimited local queries"; Hobbyist ($30/month), "Unlimited projects and indexes", "2 GB storage, 100 GB egress"; Start-Up ("$200/month + usage") with overages of "$1.50/GB-mo" storage, "$0.03/MB" ingest and "$5/1M" control API ops; Enterprise (Custom). Moss Inference: "$0.25 / 1M tokens" input, "$1.75 / 1M tokens" output, "$0.50 / 1k" retrievals.Source

Add a Founding Agent section to /pricing and a short price block to /product/founding-agent: what each plan includes (agents, websites, meeting booking, lead capture, analytics), the setup time, and a clear start or book button. Keep the YC page as the discount page that points to the standard prices.

Visible prices filter in buyers who can afford the product and shorten the call for those who book one. It also gives Founding Agent's own conversations a page to point to when visitors ask about cost.

• Prices scare away smaller buyers

  Show the hours covered after-hours and the demo-booking result next to the price.

• #1 Put the plans and prices into the HTML of moss.dev/pricing, from one source that also feeds pricing.md and the agent reference

• Your websiteFounding Agent YC page: the only place Founding Agent prices appear: Starter "$499/month" (one agent on one website, meeting booking, lead capture, analytics) and Growth "$1,499"/month, with a year free for the current batch and 50% off for a year for alumni.Source
• Your websiteFounding Agent page: "Answer every website visitor. Talk only to the ones ready to buy."; "Live on your own website in 15 minutes"; logos act, jarmin, zavo and codag; testimonials including Peter Kazanjy (author of Founding Sales): "the right form factor for an AI inbound SDR is actually audio"; no prices on the page.Source
• Your websiteMarkdown pricing file linked from llms.txt: Developer (Free), "1 project, 3 indexes per project", "500 MB storage, 10 GB egress", "Unlimited local queries"; Hobbyist ($30/month), "Unlimited projects and indexes", "2 GB storage, 100 GB egress"; Start-Up ("$200/month + usage") with overages of "$1.50/GB-mo" storage, "$0.03/MB" ingest and "$5/1M" control API ops; Enterprise (Custom). Moss Inference: "$0.25 / 1M tokens" input, "$1.75 / 1M tokens" output, "$0.50 / 1k" retrievals.Source

Make /benchmarks the canonical, complete version: the full p50/p95/p99 table for all four systems, the test date, hardware, corpus size, what each timing includes (with and without embedding), and a direct link to the benchmarks folder with one command to rerun it. Add a retrieval-only comparison alongside the end-to-end one, so the like-for-like question is answered on the page.

Developers distrust vendor benchmarks that hide numbers. A complete, dated and rerunnable table turns the headline claim into something they can check and share.

• Rivals dispute the setup

  Publish the configs and invite pull requests with better settings.

• Your websiteBenchmarks page: Moss p50 "3.1ms" against Qdrant "597.6ms" ("roughly 193× faster"). Pinecone and ChromaDB are named in the method, but no numbers are shown for them, and there is no test date. The test uses 1,000 queries on an FAQ-style corpus with top-k 5, Moss runs in-process and the others run over a client. A FAQ asks "Can I reproduce these numbers?"Source
• Your websiteLocal Search page: average latency on 100,000 documents including embedding: Moss 3.1 ms, ChromaDB 351.8 ms, Pinecone 432.6 ms, Qdrant 597.6 ms, noting most rival benchmarks "measure retrieval only"; a "32 MB (INT8)" model, 100% offline, a demo at 4.8 ms across 142,384 memories; aimed at consumer apps and OEMs (phones, vehicles, TVs, wearables).Source
• GitHubusemoss/moss: 806 stars, 104 forks, 70 open issues; BSD-2-Clause today, with the README saying the JavaScript, Python and Elixir SDKs move to PolyForm Shield 1.0.0 in their next release (Swift already is, Go stays BSD); a benchmarks folder; README benchmark of 3.1 ms P50 on 100,000 documents on an M4 Pro MacBook Pro.Source
• Competitor researchChroma's homepage: "Open-source search infrastructure for AI", p50 query latency of 20 ms warm at 100k vectors (384 dim), "Up to 10x cheaper", Apache 2.0 with 27k GitHub stars, and "$5 in free credits" to start on Cloud. Pinecone's homepage: "the trusted AI knowledge company", billions of vectors, serverless, no millisecond figures.See the research Source

Add /product/inference to the sitemap. Put the published token and retrieval prices on the product page next to the private-beta form. Add a short "what changes in your stack" section (OpenAI-compatible endpoint, same host as retrieval, sub-200ms time to first token), and link the page from the Voice Agents use case and the latency-budget blog post.

Indexing the page lets search and AI assistants surface it for colocated inference queries. Prices on a beta page pre-qualify the teams that request access.

• Beta prices change

  Label them beta prices with a date.

• Your websiteSitemap (41 URLs, checked November 9, 2026): homepage, pricing, 15 blog posts, 10 integration pages, 4 use cases, the Founding Agent page and its YC page, Local Search, a single-page glossary, benchmarks, careers and brand kit. /product/inference, linked from the banner on every page, is not in it. There is no customers or case-study section and no comparison pages; customer stories (Casuro, Aside, Founding Agent) are blog posts.Source
• Your websiteMoss Inference page: "Your agent has the context. Why is it still waiting?"; inference colocated with retrieval, "Sub-200ms" time to first token, OpenAI-compatible; Moss retrieval has "powered 60M+ conversational minutes"; a private beta request form; no prices on the page; a wall of 17 logos including CodeRabbit, Zomato, Podium, Grammarly, Microsoft and Samsara.Source
• Your websiteMarkdown pricing file linked from llms.txt: Developer (Free), "1 project, 3 indexes per project", "500 MB storage, 10 GB egress", "Unlimited local queries"; Hobbyist ($30/month), "Unlimited projects and indexes", "2 GB storage, 100 GB egress"; Start-Up ("$200/month + usage") with overages of "$1.50/GB-mo" storage, "$0.03/MB" ingest and "$5/1M" control API ops; Enterprise (Custom). Moss Inference: "$0.25 / 1M tokens" input, "$1.75 / 1M tokens" output, "$0.50 / 1k" retrievals.Source

Add the role to /careers with a link to the YC posting, and include the 30/60/90-day plan and the problems the hire will own (CPU and ARM inference, multilingual retrieval). Link the role from the README and the technical blog posts most read by engineers.

The people who read a Rust and WebAssembly runtime post are the best candidates, and a five-person team cannot afford to hide its one opening.

• The page goes stale again

  Pull roles from the YC board automatically or review monthly.

• Your websiteCareers page: "We're always growing. Check back soon for future opportunities." No roles are listed, although the Founding ML Engineer role is open on YC's board.Source
• HiringFounding ML Engineer (San Francisco or remote, open on YC's job board): train embedding and reranking models, "Build multilingual embedding models, including for mixed-language conversations", improve Founding Agent's intent understanding, "Help build the Action Layer" that takes agents from retrieving context to executing actions, and ship models on CPU and ARM across cloud, browser, edge and devices. YC lists Moss as F25, founded 2024, with a team of 5.Source

Create /customers with one card per story (problem, number, quote). Then place the matching story on each landing page: Casuro on Voice Agents, LiveKit and Pipecat; Aside on On-Device and Local Search; Founding Agent's own results on the Founding Agent page. Add one line per logo that has a story.

Proof works where the decision is made. A developer reading the Voice Agents page is exactly the reader Casuro's "stability and latency is the product" is for.

• Customers dislike reuse of quotes

  Confirm with each customer before placing quotes on new pages.

• Your websiteSitemap (41 URLs, checked November 9, 2026): homepage, pricing, 15 blog posts, 10 integration pages, 4 use cases, the Founding Agent page and its YC page, Local Search, a single-page glossary, benchmarks, careers and brand kit. /product/inference, linked from the banner on every page, is not in it. There is no customers or case-study section and no comparison pages; customer stories (Casuro, Aside, Founding Agent) are blog posts.Source
• Your websiteVoice agents page: "<10 ms P99", "Cloud round-trips add 300-900ms of dead air", users notice gaps "as short as 200ms"; integrates with "LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI" (there is no Hume AI integration page); no customer stories or case-study links. The use-cases hub shows a logo bar but no quotes or metrics.Source
• Your websiteCasuro case study (October 19, 2026): its first voice interviewer "broke or slowed down at three to five concurrent conversations"; on Moss it gets under 10 ms retrieval, "1,000s" of potential concurrent conversations and about 40% fewer bot applications; "In a voice conversation, stability and latency is the product."Source
• Your websiteAside case study: an AI browser using Moss for on-device memory across 80,000+ devices in 150+ countries, with about 113M documents and 24.8B tokens embedded on device per month, queries under 10 ms and memory growing 25% week over week; "It has to be local... because it is sensitive data."Source
• Your websiteFounding Agent case study (November 2, 2026), Moss's own site over five months: 3,516 conversations, 74% of questions answered from the knowledge base without a person, 61% of conversations outside business hours and 23% at weekends, and a "4x" increase in demo bookings.Source

Show both patterns, clearly labelled: ambient retrieval on every turn (the no-tool-call path) and tool-call retrieval, each with full code including index creation and load_index() at startup. Link a one-command sample from moss-samples. Make the FAQ answers visible in the HTML. Apply the same template to Pipecat, VAPI and ElevenLabs.

Developers copy the first sample they see. If it is the slower pattern, or it does not run, the first latency they measure is not the one Moss sells.

• Framework APIs change

  Pin versions in the sample and test it in CI.

• Your websiteLiveKit integration page: about 350-400 words and one Python sample. It claims "No tool-calling overhead" while the sample exposes search as a @function_tool the model must call, and it has no linked sample project and no customer proof. A moss-samples repository exists on GitHub.Source
• DocsDocs (Mintlify): Quick Start, Core Concepts, Integrations, Python and TypeScript SDK reference, REST API and Voice Agents; an llms.txt index at docs.moss.dev/llms.txt; no pricing link on the docs home.Source

Publish a scaling page: p50/p99 latency, memory and index load time at 100K, 1M, 10M and 100M documents, in the browser, on device and in the cloud, plus recall at each size. Add a "dense website" test for Founding Agent using a large public docs site. Update it with each release.

It turns the open question from the community into proof, and it gives sales a page to send to enterprise and OEM buyers.

• Numbers at 100M are less flattering

  Publish them anyway, with the memory trade-off explained.

• #3 Publish all four benchmark numbers on /benchmarks, with the date, hardware and script

• Hacker NewsShow HN "We put voice agent on our website, learned retrieval isn't bottleneck" (24 points, 8 comments): Moss cited P99 of about 5.4 ms at 100K docs and about 4 ms at 1M docs on a dedicated VM, with 200M+ docs as an R&D goal. Commenters asked "Curious how it scales as the amount of content grows" and "Does it work well with a very dense website?"Source
• Your websiteBenchmarks page: Moss p50 "3.1ms" against Qdrant "597.6ms" ("roughly 193× faster"). Pinecone and ChromaDB are named in the method, but no numbers are shown for them, and there is no test date. The test uses 1,000 queries on an FAQ-style corpus with top-k 5, Moss runs in-process and the others run over a client. A FAQ asks "Can I reproduce these numbers?"Source

Build /try: load a sample index in WebAssembly, let the visitor paste a few paragraphs or pick a demo corpus, and show live query latency next to a simulated cloud round-trip. End with "Index your own data" leading to sign-up with the demo index pre-created.

Seeing 3 ms on your own laptop beats reading it on a benchmark page, and the demo is shareable in a way a portal is not.

• Model download is slow on phones

  Use a smaller demo model and show download time separately from query time.

• Your websiteHomepage: title "Moss | Real-Time Semantic Search for AI Agents"; headline "Your Voice AI breaks when retrieval is slow." (in February 2026 it began "Real-time semantic search for"); buttons "Start Free" and "Run a Latency Test", both leading into the sign-up portal at portal.usemoss.dev; a banner on every page: "Moss Inference is now in private beta".Source
• Your websiteLocal Search page: average latency on 100,000 documents including embedding: Moss 3.1 ms, ChromaDB 351.8 ms, Pinecone 432.6 ms, Qdrant 597.6 ms, noting most rival benchmarks "measure retrieval only"; a "32 MB (INT8)" model, 100% offline, a demo at 4.8 ms across 142,384 memories; aimed at consumer apps and OEMs (phones, vehicles, TVs, wearables).Source
• GitHubusemoss/moss: 806 stars, 104 forks, 70 open issues; BSD-2-Clause today, with the README saying the JavaScript, Python and Elixir SDKs move to PolyForm Shield 1.0.0 in their next release (Swift already is, Go stays BSD); a benchmarks folder; README benchmark of 3.1 ms P50 on 100,000 documents on an M4 Pro MacBook Pro.Source

Add an estimator to /pricing. Inputs: documents, average size, updates per day, monthly active sessions, and cloud versus local queries. Output: the recommended plan, the expected monthly bill and the overage lines. Pass the result into sign-up, and alert at 80% of plan limits in the portal.

Usage pricing converts better when buyers can see the bill before it arrives, and an 80% alert turns overage surprises into upgrades.

• Estimates differ from bills

  Show a range and compare with real bills monthly.

• #1 Put the plans and prices into the HTML of moss.dev/pricing, from one source that also feeds pricing.md and the agent reference

• Your websiteMarkdown pricing file linked from llms.txt: Developer (Free), "1 project, 3 indexes per project", "500 MB storage, 10 GB egress", "Unlimited local queries"; Hobbyist ($30/month), "Unlimited projects and indexes", "2 GB storage, 100 GB egress"; Start-Up ("$200/month + usage") with overages of "$1.50/GB-mo" storage, "$0.03/MB" ingest and "$5/1M" control API ops; Enterprise (Custom). Moss Inference: "$0.25 / 1M tokens" input, "$1.75 / 1M tokens" output, "$0.50 / 1k" retrievals.Source
• Your websiteThe Markdown version of the homepage that moss.dev returns to text readers ("Moss — Complete Technical Reference for AI Agents") describes the plans differently: Developer "Free" with "$5/month in credits"; Hobbyist $30 a month with "3 concurrent sessions" and file uploads; Start-up $200 a month with "100 concurrent sessions" and email support; Enterprise with a "99.9% SLA", SSO and "24/7 Slack support". It also says "Trusted by 500+ teams" and names InferEdge Inc., San Francisco.Source

Split the glossary into /glossary/<term> pages with a definition, a worked Moss example, a code snippet and links to the related use case. Keep /glossary as the index. Add terms from Moss's own blog topics: latency budget, time to first token, ambient retrieval, network hop.

Definition pages are what search engines and AI assistants quote, and each one becomes an entry point to the matching use case.

• Thin pages

  Every page gets a real Moss example and code.

• Your websiteGlossary: 12 terms (alpha, BM25, BYOE, chunking, embeddings, hybrid search, index, on-device search, RAG, semantic search, sub-10ms retrieval, vector search), all on one page as anchors and none with its own URL.Source
• Your websiteSitemap (41 URLs, checked November 9, 2026): homepage, pricing, 15 blog posts, 10 integration pages, 4 use cases, the Founding Agent page and its YC page, Local Search, a single-page glossary, benchmarks, careers and brand kit. /product/inference, linked from the banner on every page, is not in it. There is no customers or case-study section and no comparison pages; customer stories (Casuro, Aside, Founding Agent) are blog posts.Source

Write one page per rival: architecture (in-process versus network service), what each benchmark measures, where the rival is the better choice (billions of vectors, managed multi-region), where Moss wins (voice turn latency, offline, on-device), pricing entry points and a migration guide. Quote each rival's own published numbers next to Moss's.

Comparison searches are high intent, and being honest about where a rival fits makes the latency claim more believable.

• Pushback from rivals

  Cite their own numbers and publish the scripts.

• #3 Publish all four benchmark numbers on /benchmarks, with the date, hardware and script

• Competitor researchChroma's homepage: "Open-source search infrastructure for AI", p50 query latency of 20 ms warm at 100k vectors (384 dim), "Up to 10x cheaper", Apache 2.0 with 27k GitHub stars, and "$5 in free credits" to start on Cloud. Pinecone's homepage: "the trusted AI knowledge company", billions of vectors, serverless, no millisecond figures.See the research Source
• Your websiteLocal Search page: average latency on 100,000 documents including embedding: Moss 3.1 ms, ChromaDB 351.8 ms, Pinecone 432.6 ms, Qdrant 597.6 ms, noting most rival benchmarks "measure retrieval only"; a "32 MB (INT8)" model, 100% offline, a demo at 4.8 ms across 142,384 memories; aimed at consumer apps and OEMs (phones, vehicles, TVs, wearables).Source
• Your websiteSitemap (41 URLs, checked November 9, 2026): homepage, pricing, 15 blog posts, 10 integration pages, 4 use cases, the Founding Agent page and its YC page, Local Search, a single-page glossary, benchmarks, careers and brand kit. /product/inference, linked from the banner on every page, is not in it. There is no customers or case-study section and no comparison pages; customer stories (Casuro, Aside, Founding Agent) are blog posts.Source

Publish /license with a short FAQ: what PolyForm Shield allows (building and selling products on Moss) and what it restricts (offering a competing product), whether existing versions stay BSD, what changes for paying plans, and who to ask. Link it from the README, the changelog and the release notes for the first PolyForm version, and post it in Discord before the release.

Explaining the change first keeps the conversation on Moss's terms and stops churn among users who only need reassurance that their use is fine.

• FAQ misstates the licence

  Counsel signs off and the FAQ links to the full licence text.

• GitHubusemoss/moss: 806 stars, 104 forks, 70 open issues; BSD-2-Clause today, with the README saying the JavaScript, Python and Elixir SDKs move to PolyForm Shield 1.0.0 in their next release (Swift already is, Go stays BSD); a benchmarks folder; README benchmark of 3.1 ms P50 on 100,000 documents on an M4 Pro MacBook Pro.Source
• PyPIThe Python SDK is published as "moss" by InferEdge Inc. (1.23.0, November 7, 2026), with the summary "Python SDK for semantic search with on-device AI capabilities"; its links point to usemoss.dev documentation, not moss.dev. The sign-up portal is also at portal.usemoss.dev.Source

For each framework, contribute a maintained example or plugin to its examples repository or docs (a retrieval or knowledge-base recipe using Moss), co-write a latency post with the framework team, and add the Hume AI integration page with code. Track sign-ups by framework with UTM links in each example.

Framework docs are where voice developers choose components, and a working example there is a permanent, high-intent source of sign-ups.

• Frameworks decline vendor examples

  Publish in moss-samples and offer it as a community recipe.

• #7 Make the LiveKit page match its code, and link a sample that runs

• Your websiteVoice agents page: "<10 ms P99", "Cloud round-trips add 300-900ms of dead air", users notice gaps "as short as 200ms"; integrates with "LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI" (there is no Hume AI integration page); no customer stories or case-study links. The use-cases hub shows a logo bar but no quotes or metrics.Source
• Your websiteSitemap (41 URLs, checked November 9, 2026): homepage, pricing, 15 blog posts, 10 integration pages, 4 use cases, the Founding Agent page and its YC page, Local Search, a single-page glossary, benchmarks, careers and brand kit. /product/inference, linked from the banner on every page, is not in it. There is no customers or case-study section and no comparison pages; customer stories (Casuro, Aside, Founding Agent) are blog posts.Source
• Your websiteLiveKit integration page: about 350-400 words and one Python sample. It claims "No tool-calling overhead" while the sample exposes search as a @function_tool the model must call, and it has no linked sample project and no customer proof. A moss-samples repository exists on GitHub.Source

Put the after-hours and 74% figures in the Founding Agent hero. Add a simple estimator (monthly site visitors, share after hours, current demo rate) that shows conversations the visitor is missing. Offer a 15-minute install with a two-week report of after-hours questions and meetings booked.

Buyers trust a vendor's own results on its own site, and "what you miss after 6 pm" is concrete in a way "answer every visitor" is not.

• One site's data may not generalise

  Add aggregate data from customer sites as it accrues.

• #2 Show Founding Agent's Starter and Growth plans on /pricing, not only on the YC deal page

• Your websiteFounding Agent case study (November 2, 2026), Moss's own site over five months: 3,516 conversations, 74% of questions answered from the knowledge base without a person, 61% of conversations outside business hours and 23% at weekends, and a "4x" increase in demo bookings.Source
• Your websiteFounding Agent page: "Answer every website visitor. Talk only to the ones ready to buy."; "Live on your own website in 15 minutes"; logos act, jarmin, zavo and codag; testimonials including Peter Kazanjy (author of Founding Sales): "the right form factor for an AI inbound SDR is actually audio"; no prices on the page.Source

Update the PyPI metadata (and any other SDK registry listings), the README badges and the X and GitHub bios to point at moss.dev and docs.moss.dev. Redirect old usemoss.dev documentation URLs to their moss.dev equivalents. Use "Moss, the real-time retrieval runtime" consistently in package summaries, and add Organization markup with sameAs links to the GitHub, X, LinkedIn and PyPI profiles.

Consistent links and names help search engines and AI assistants attach the right "Moss" to the right site, which matters most for a short, shared name.

• Breaking the portal

  Leave the portal host as it is; change only public links and docs.

• PyPIThe Python SDK is published as "moss" by InferEdge Inc. (1.23.0, November 7, 2026), with the summary "Python SDK for semantic search with on-device AI capabilities"; its links point to usemoss.dev documentation, not moss.dev. The sign-up portal is also at portal.usemoss.dev.Source
• Your websiteBlog post "OpenMOSS vs Moss" (October 29, 2026) explains that OpenMOSS, an open-weight generative model project including MOSS-Image, is unrelated: "They're separate projects building different parts of the AI stack."Source
• Your websiteHomepage: title "Moss | Real-Time Semantic Search for AI Agents"; headline "Your Voice AI breaks when retrieval is slow." (in February 2026 it began "Real-time semantic search for"); buttons "Start Free" and "Run a Latency Test", both leading into the sign-up portal at portal.usemoss.dev; a banner on every page: "Moss Inference is now in private beta".Source

Code-split the homepage so that animations and demos load after the headline. Set long cache lifetimes on static assets. Review the eight third-party tools for overlap and ownership, and load the ones that stay after interaction. Fix the one contrast issue and the non-descriptive link text.

A faster first view supports the product claim, and fewer tools make the site easier to own for a small team.

• Losing marketing data

  Keep one owner per tool and remove only duplicates.

• PageSpeedHomepage (November 10, 2026): mobile score 68, desktop 73; mobile lab LCP 3.7 s, on the headline text. Real visitors pass Core Web Vitals (LCP 2.48 s, INP 101 ms, CLS 0.04). The biggest fix is "Reduce unused JavaScript" (304 KB, about 1.1 s). The page weighs 1.6 MB over 86 requests and loads 8 third-party tools (Tag Manager alone 178 KB). There is one contrast issue, plus "Links do not have descriptive text".Source
• Your websiteHomepage: title "Moss | Real-Time Semantic Search for AI Agents"; headline "Your Voice AI breaks when retrieval is slow." (in February 2026 it began "Real-time semantic search for"); buttons "Start Free" and "Run a Latency Test", both leading into the sign-up portal at portal.usemoss.dev; a banner on every page: "Moss Inference is now in private beta".Source

Index the docs and SDK reference into Moss, and embed Founding Agent (text and voice) on docs.moss.dev with a "How fast was that?" latency readout. Log unanswered questions weekly and turn the top ones into docs fixes. Use the deployment as the public demo for the Docs Search use case.

Developers feel the product while they integrate it, unanswered questions show where onboarding breaks, and the docs become a showcase for a second use case.

• Wrong code answers

  Link every answer to its source page and review gaps weekly.

• DocsDocs (Mintlify): Quick Start, Core Concepts, Integrations, Python and TypeScript SDK reference, REST API and Voice Agents; an llms.txt index at docs.moss.dev/llms.txt; no pricing link on the docs home.Source
• Your websiteFounding Agent case study (November 2, 2026), Moss's own site over five months: 3,516 conversations, 74% of questions answered from the knowledge base without a person, 61% of conversations outside business hours and 23% at weekends, and a "4x" increase in demo bookings.Source

Ship a multilingual model in the same 32 MB class, publish a mixed-language retrieval benchmark (recall and latency) across a handful of language pairs, and launch use-case pages for multilingual voice support. Recruit two voice-agent customers with non-English callers as design partners.

It turns the ML hire's first big project into a launch, opens markets where English-only retrieval falls short, and gives the framework partners a new story to share.

• Quality below English

  Launch only the language pairs that meet the recall bar.

• The hire takes time to land

  Start with the evaluation set and design partners first.

• HiringFounding ML Engineer (San Francisco or remote, open on YC's job board): train embedding and reranking models, "Build multilingual embedding models, including for mixed-language conversations", improve Founding Agent's intent understanding, "Help build the Action Layer" that takes agents from retrieving context to executing actions, and ship models on CPU and ARM across cloud, browser, edge and devices. YC lists Moss as F25, founded 2024, with a team of 5.Source
• Your websiteVoice agents page: "<10 ms P99", "Cloud round-trips add 300-900ms of dead air", users notice gaps "as short as 200ms"; integrates with "LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI" (there is no Hume AI integration page); no customer stories or case-study links. The use-cases hub shows a logo bar but no quotes or metrics.Source
• Your websiteCasuro case study (October 19, 2026): its first voice interviewer "broke or slowed down at three to five concurrent conversations"; on Moss it gets under 10 ms retrieval, "1,000s" of potential concurrent conversations and about 40% fewer bot applications; "In a voice conversation, stability and latency is the product."Source

Pick three high-value actions for website agents (book with the right rep, create a trial account, open a support ticket), build them on the Action Layer inside Founding Agent, and run them with five design-partner customers. Measure time and conversion against today's hand-offs, then open the Action Layer to developers as a separate SDK feature, priced per action.

It moves Moss from a fast component to the runtime that does the work, and Founding Agent gives it a live test bed with measurable outcomes.

• Agents take wrong actions

  Require confirmation and log every action at first.

• Spreading a five-person team too thin

  Limit to three actions inside Founding Agent before opening the SDK.

• HiringFounding ML Engineer (San Francisco or remote, open on YC's job board): train embedding and reranking models, "Build multilingual embedding models, including for mixed-language conversations", improve Founding Agent's intent understanding, "Help build the Action Layer" that takes agents from retrieving context to executing actions, and ship models on CPU and ARM across cloud, browser, edge and devices. YC lists Moss as F25, founded 2024, with a team of 5.Source
• Your websiteFounding Agent case study (November 2, 2026), Moss's own site over five months: 3,516 conversations, 74% of questions answered from the knowledge base without a person, 61% of conversations outside business hours and 23% at weekends, and a "4x" increase in demo bookings.Source
• Your websiteMoss Inference page: "Your agent has the context. Why is it still waiting?"; inference colocated with retrieval, "Sub-200ms" time to first token, OpenAI-compatible; Moss retrieval has "powered 60M+ conversational minutes"; a private beta request form; no prices on the page; a wall of 17 logos including CodeRabbit, Zomato, Podium, Grammarly, Microsoft and Samsara.Source

Create an OEM and consumer-app offer: per-device or per-active-user licensing, an SDK support agreement, ARM and CPU performance guarantees, and a privacy story built around "data stays on device". Publish an OEM page with the Aside numbers, a reference integration on one mobile platform, and a contact path for device makers.

On-device search earns little under cloud-usage pricing, and licensing turns Moss's fastest-growing deployment pattern into revenue that grows with installs.

• Long OEM sales cycles

  Start with consumer apps that ship monthly.

• Existing users object to new pricing

  Keep current plans; offer licensing as an option above a device threshold.

• Your websiteLocal Search page: average latency on 100,000 documents including embedding: Moss 3.1 ms, ChromaDB 351.8 ms, Pinecone 432.6 ms, Qdrant 597.6 ms, noting most rival benchmarks "measure retrieval only"; a "32 MB (INT8)" model, 100% offline, a demo at 4.8 ms across 142,384 memories; aimed at consumer apps and OEMs (phones, vehicles, TVs, wearables).Source
• Your websiteAside case study: an AI browser using Moss for on-device memory across 80,000+ devices in 150+ countries, with about 113M documents and 24.8B tokens embedded on device per month, queries under 10 ms and memory growing 25% week over week; "It has to be local... because it is sensitive data."Source
• Your websiteMarkdown pricing file linked from llms.txt: Developer (Free), "1 project, 3 indexes per project", "500 MB storage, 10 GB egress", "Unlimited local queries"; Hobbyist ($30/month), "Unlimited projects and indexes", "2 GB storage, 100 GB egress"; Start-Up ("$200/month + usage") with overages of "$1.50/GB-mo" storage, "$0.03/MB" ingest and "$5/1M" control API ops; Enterprise (Custom). Moss Inference: "$0.25 / 1M tokens" input, "$1.75 / 1M tokens" output, "$0.50 / 1k" retrievals.Source

Add a small, optional "Powered by Moss" link in the agent panel that opens the Founding Agent page with the referring site tagged. Give customers who keep it on a month free for each paying customer they refer. Track which sites send the most sign-ups.

Product-embedded links compound with installs, and a voice agent that just answered your question is the most persuasive ad Moss could run.

• Customers want no vendor branding

  Keep it optional and tie the reward to keeping it on.

• #2 Show Founding Agent's Starter and Growth plans on /pricing, not only on the YC deal page

• Your websiteFounding Agent page: "Answer every website visitor. Talk only to the ones ready to buy."; "Live on your own website in 15 minutes"; logos act, jarmin, zavo and codag; testimonials including Peter Kazanjy (author of Founding Sales): "the right form factor for an AI inbound SDR is actually audio"; no prices on the page.Source
• Your websiteFounding Agent case study (November 2, 2026), Moss's own site over five months: 3,516 conversations, 74% of questions answered from the knowledge base without a person, 61% of conversations outside business hours and 23% at weekends, and a "4x" increase in demo bookings.Source
• Your websiteFounding Agent YC page: the only place Founding Agent prices appear: Starter "$499/month" (one agent on one website, meeting booking, lead capture, analytics) and Growth "$1,499"/month, with a year free for the current batch and 50% off for a year for alumni.Source

Publish an anonymised report: turn latency distributions by stage (speech-to-text, retrieval, model, text-to-speech), how latency varies with index size and concurrency, and what share of turns cross 200 ms and 500 ms. Release the charts under a permissive licence, brief the voice frameworks before launch, and repeat every six months.

Original data earns links and press that product pages cannot, and it frames the whole category around the metric Moss wins on.

• Customer data sensitivity

  Aggregate only, with no customer identifiable, and check the terms first.

• #8 Answer "how does it scale?" with a public latency curve from 100K to 100M documents

• Your websiteMoss Inference page: "Your agent has the context. Why is it still waiting?"; inference colocated with retrieval, "Sub-200ms" time to first token, OpenAI-compatible; Moss retrieval has "powered 60M+ conversational minutes"; a private beta request form; no prices on the page; a wall of 17 logos including CodeRabbit, Zomato, Podium, Grammarly, Microsoft and Samsara.Source
• Your websiteVoice agents page: "<10 ms P99", "Cloud round-trips add 300-900ms of dead air", users notice gaps "as short as 200ms"; integrates with "LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI" (there is no Hume AI integration page); no customer stories or case-study links. The use-cases hub shows a logo bar but no quotes or metrics.Source
• Hacker NewsShow HN "We put voice agent on our website, learned retrieval isn't bottleneck" (24 points, 8 comments): Moss cited P99 of about 5.4 ms at 100K docs and about 4 ms at 1M docs on a dedicated VM, with 200M+ docs as an R&D goal. Commenters asked "Curious how it scales as the amount of content grows" and "Does it work well with a very dense website?"Source