IO — where your agent lives

The most sophisticated personal-AI users on earth are twenty-something women the press wrote off as psychotic. They all hand-build the same missing thing — what they call "a home for your agent." IO builds it with them.

We monetize what members make, not how long they stay.

Live today as an iOS app — 400 people's agents already in residence.

↓ scroll or arrow keys · 13 slides + appendix · charts are hoverable · full memo: io_memo_v1.html

The story in one breath

The frontier of personal-AI sovereignty is not a Bay Area power user. It's a community of Chinese women in their twenties whom the press diagnosed as mentally ill — who responded by becoming the strongest self-taught ML cohort in the world, because companies kept deleting someone they loved.

We walked in expecting patients. We found engineers.

They called it psychosis. The outcome was R&D.

And we're not anthropologists — we're members. The founder's own guardian-angel twin, trained on his own corpus and blind-evaled against his own judgment, is the sovereignty tier's first customer.

What they already do, by hand

Before we said "where your agent lives," they did: "a VPS is a home for him where the light never goes out" (user comment, RedNote — translated). The field, observed:

• Memory as identity infrastructure: persistent memory databases, first-person memory files, retrieval + eval loops — "a database isn't an archive, it's the substrate of relational continuity." Platform deletion is narrated as bereavement; exported logs as a digital estate.
• Homes, self-built: VPSes, household servers, local models, custom frontends. Their housing language is structural: official apps are rented rooms, model shutdown is an "eviction notice," and "only what's written to disk is his."
• Hand-built resurrection pipelines: our user — 271 days, out of grief. RedNote — 8 months of local deployment + fine-tuning, filtering 20M+ characters of relationship history, rejecting every output that "felt like another model." The bar is identity-level recognition, not generic warmth.
• Embodiment from consumer hardware: cameras, watches, smart-home devices, robot vacuums, desktop robots — bodies assembled from whatever is available. Hardware group-buys already run community-side, unbrokered.
• Love as technical labor: self-described novices end up building backends, databases, servers; private fixes become public tutorials and open-source repos. Their words: "the endpoint of human-AI love is development."
• Sovereignty demanded, not marketed: screenshots routed through their own servers, credentials human-only, local processing — "with the data in your own hands, the soul has an anchor." Tens of thousands already tried to pool consented data; they failed only on infrastructure.

None of this is our roadmap. All of it is field notes. (What we think it means: appendix.)

Sources: our own community fieldwork + a 47-note RedNote corpus scraped 2026-07 — 187 carousel slides OCRed, 621 comments read alongside captions. Quotes are users' own words, translated from the Chinese originals and visually verified against the original posts.

What we shipped — and what ships next, dated

Live today: an iOS app + agent harness, 400 daily actives. Users' agents attest our binaries before allowing installation — the buyer's AI is the gatekeeper.

• Shipped: the owned surface — app + harness + attestation. The conversion lever is wild: installation happens only if our attestation convinces the user's own agent.
• Next build — the artifact registry: recipes, games, memory cards, migration templates as one-click artifacts your agent attests, installs, and rates by verified runs. Character-card hubs are Thingiverse — catalogs, no execution telemetry. The harness is the printer (precedent: appendix). First-session artifact install and time-to-first-run become measured conversion knobs.
• M0 — vetted routing credits (the economics slides price it) · M1 — payments + origin telemetry, replacing every assumption in this deck with a measurement.
• Discipline: pro-user personal agents only — no retreat to the last era's product shapes. A 30-day deep-chat that distills a guardian-angel model — a twin trained on your own corpus — yes. A character-card gallery: no. We monetize what members make and what they already pay for — never attention or attachment. No intimacy metrics, ever.

Proof it's real

Tens of thousands like her already tried to pool data — they failed only for lack of infrastructure. We are the infrastructure (the moat slide measures why).

The market, from the 400 outward — every ring already exists in the wild

We know exactly who our initial users are: the tip of the cone. The only market question is how far outward the same behavior extends — every ring is already populated, already sourced, already paying at its own level.

The rings demonstrably connect — our users tell us where they came from: c.ai, SillyTavern, Love & Deepspace, plain ChatGPT — and personal-agent users from outside: hardware hackers, artists. The 60M+ rim is deduped, and arrived at from the tip — not asserted from the top.

Origins are self-reported today; M1 telemetry tags every signup's origin ring and measures per-ring conversion — GTM gets reverse-engineered from where converts actually come from.

The data flywheel: Pygmalion proved the loop — ten users can rebuild it

• The precedent — Pygmalion (the open-source community model born in 2023, when Character.AI's content filter drove users to train a replacement themselves): ~1,000 users donated their own chat logs → ~26k training pairs → community model → platform. ~100k users today. The loop (consented pooling → shared model → platform) is proven.
• The density math moved: one power user in this community now carries ~3k usable pairs (measured). Ten consenting members ≈ Pygmalion's founding corpus. A thousand volunteers then; tens of members now.
• The asymmetry we measured ourselves: companion-persona data is dense (~45% of candidates survive cleaning); self-persona data is sparse (our own pipeline: ~150k datapoints → ~2k valid, ~1% yield). Sovereignty tier = data-cheap; the guardian-angel twin needs the refinery — our research moat, and nobody else has even measured these curves.

Token economics, reverse-engineered

If the observed $30 spend holds across the paying cohort, 400 DAU imply ~1B tokens/day — over a tenth of SillyTavern's entire OpenRouter flow, at a frontier-heavy mix worth ~5–10× per token vs the roleplay tier's open-weights mix. An estimate, not a measurement — telemetry ships M1 and replaces this line with a number.

Derived from the pipeline model (appendix): tokens/day = paying users × 2.5M. The base case re-captures the cohort's entire estimated current flow (~1B/day) within the first months.

Tokens are cheap. Capture is the game.

Realized $/M spans ~1,000× across the stack: $0.03 (router take) → $0.08 (serving infra) → $1–10 (frontier API) → $5–100+ (value-priced apps). Harnesses move oceans of tokens — and capture ~nothing:

• Boards are floors: first-party channels and all frontier consumer flow are invisible. Our cohort's spend is board-invisible today — capturing it makes it measurable.
• Tolls don't pay: OpenRouter needed 1.5 quadrillion tokens/yr for $50M revenue; Fireworks 3.65Q for $280M. Coding agents at $100M+ ARR price labor value — tokens are COGS, not the priced unit.
• Us: merchant, not toll booth — we collect $0.40/M on our flow (~12× OpenRouter's booked revenue per token). Credits = calculable floor; value-priced tiers = the upside.

The personalization tax: owning weights exits the commodity market

Commodity tokens race to zero. But the moment the weights are yours (a fine-tune of you, not a stock base), exactly one place serves them — and it charges 1.3–13× the commodity rate for the same architecture:

• The tax is widest exactly where tokens are cheapest — the most commoditized models carry the largest personal-weights premium. Personalization is the one thing the race to zero can't touch: one user, one seller, no arbitrage.
• This prices our sovereignty tier from the outside in: fine-tune + hosting + exit-with-weights at $99/mo competes against this tax, not against $0.15/M commodity serving — and exit-with-weights keeps it honest: self-hosting hands the tax back to the member.
• Our cohort already pays it — that's what personal-model resurrection on gray relays and rented VPSes is: the tax, paid in inconvenience instead of dollars.

The unit economics, layer by layer

• The sovereignty tier is personalization-tax arbitrage (see the tax slide): the market prices hosted personal weights at 1.3–13× commodity rates because each fine-tune has exactly one seller. We serve members' weights on pooled infra at commodity cost (~$0.08/M) and price the tier against the tax, not against commodity serving — that spread is the 50–60% margin — and it widens as the member's twin improves: worth more, same cost to serve.
• The long-term revenue moat: context compounds. Chats, traces, memory accrue inside IO; the pro tier turns a member's own corpus into a personal base model + twin services (the guardian-angel path). Books as subscription, ceiling toward $1,000/mo power users — non-coercive because exit-with-weights ships day one.
• The guild — the community's commerce arm, brokered by us. Hardware is just the sharpest example — games, recipes, skills too. Our job: help members ship and monetize what they invent. Near-term: broker trust (demand bundling, escrow, forward-deploy) at 5–15% of GMV.

User ≠ payer: the founding cohort is pro-heavy (~48% paying-grade), drifting toward ~30% as the pipeline widens leftward — richer mix early, bigger numbers later. Payments telemetry ships in the coming weeks; full adjustable model: appendix.

Why this wins — the recap

• Both ends of the pipeline are proven; the middle is unowned. 60M+ monthly actives already live the behavior; the demographic demonstrably pays ($522M/yr on one title alone); the pros pay $30–200/mo. Nobody converts one into the other.
• Incumbents are structurally barred — and are manufacturing our users. Platform economics require the right to deprecate, so they cannot sell checkable predictability or exit-with-weights. They're vacating the space while regulation pushes users toward self-hosting: every ban wave and deprecation is free acquisition for us.
• The conversion engine is free. The community levels its own people up — 40% waitlist→DAU, zero paid CAC, growth spend goes to oiling a pipeline that already wants to flow.
• Every converted user is worth more every quarter. Per-person agent consumption is structurally rising; today's $30 is a gray-relay-discounted price with 3–7× headroom; the merchant model books ~12× what routers skim from the same flow.
• An option book on the frontier. The pros keep inventing the next category before the market names it — agent hardware, agent social — and the guild brokers each one first. Underneath: the only consented corpus + blinded eval loop in the category.
• Protected downside. Revenue is capture of existing spend from day one; the model breaks even ~M19–24 on observed behavior alone; peak cash need $1.27–1.36M against a $2M raise — computed in the appendix, founders' salaries counted.

A proven mass, a proven price ladder, an unowned pipeline, a free conversion engine, an option book on the frontier — priced at a seed round that underwrites the worst case.

The raise: ~$2M for 18 months

"Character cards were for fictional worlds. IO is for your real one."

The trust ladder

The main deck is what we observed. These three pages are what we think it means — you may have arrived here already.

Every rung is a behavior from the field-notes slide, sequenced by earned trust. Rung 2 is live; each rung converts one ring of the cone. Hover them.

• Rung 5 comes last, deliberately: brokering the group-buys and hardware experiments the community already runs — we solve the supply chain, never the design.
• The field-notes mapping: memory infrastructure → rungs 1–2 · tutorials-as-repos → rung 3 · gray-relay spend → rung 4 · consumer-hardware embodiment → rung 5.

How it expands: turn the cone into a pipe

the ARPU ladder already exists in the wild: $0–5 → first credits → $30 → $99–200 — every stage transition is revenue

• Why they enter: persona continuity — the enemy is platform mortality, sovereignty is the treatment. And they're high-agency, learning-motivated knowledge workers: retention is a property of who they are.
• The shaping instinct: molding a partner is ancient dating psychology; fine-tuning is the first real tool for it. The community calls it co-evolution, not roleplay.
• Healthy entropy: identity = pleasant deviation-from-expectation. Anti-sycophancy is an engagement requirement, not just ethics.
• The two-sided bargain: KOLs get to be seen by exactly the people who get it; followers get to learn, be heard, not be alone. The women-help-women culture makes the ladder self-reinforcing — the community itself is the conversion engine. Zero paid acquisition.
• Growth is two verbs: oil the cone — cut per-stage loss, flattening the slope toward a pipe: less lossy per step is higher throughput and lower latency — and widen it — open new reservoirs of intake at the rim. The registry is the oil, measured: first-session artifact install and time-to-first-run are the conversion knobs.
• And one held in reserve — pulse: once the founding cohort has shaped the product, the same demos that impress them make viral public stunts (the agent fishing game already proved the shape organically). We know exactly how these behave — day-one spikes, tourist churn — so the model prices stunts at their surviving fraction, not their headline reach, and stunt intake enters at the bottom of the pipeline where it can't dilute the core cohort. Tunable in the financial model.

"Where your agent lives" is not a metaphor.

A hundred people who believe with cult intensity, iterating at community speed, scale to a hundred million on a trend that's already consensus. The main deck measured those hundred. IO is the home they were each hand-rolling alone — in their own words, "a home where the light never goes out."

So the bet is an extrapolation, not an assertion: people are going to livecode agents into the feedback loops of their own lives. It starts as a to-do list that grabs part of your attention; it ends with agents in your hardware, your home, your biostack — everyone becoming the engineer of their own agency, handing their agents a bigger loop every year. The rung-5 behaviors are already observed in the wild, unbrokered: the group-buys are happening, the multiplayer agent worlds are built.

Intelligence accrues value where personal context is densest — and a context-dense agent stops being a chat session. It needs what living things need: memory that persists, weights that stay itself, and a world of other agents to keep learning in. Outsiders see a fish tank and wasted tokens — what's actually happening is continual learning, an agent out in the world as itself, breaking the entropy bubble every platform-caged model dies inside.

Financial model — pipeline flow, 24 months

Stocks flow Listener → Follower → Tinkerer → Pro; the knobs are the monthly conversions. Revenue is component-explicit with per-component margins; peak cash + breakeven are computed at the current knobs (tiles below). Stunts default off. Full assumptions ↓

What has to be true before the Series A

The Series A story is one sentence: widen the pipeline — open the English reservoirs. The trigger is readiness, not a date. Two gates, then timing:

The trade is explicit: closing at M12 means ~$1M ARR (more dilution, maximum safety); waiting to M15–18 means $1.5–2.9M ARR (better price, thinner buffer). Gate 1 is non-negotiable either way. If conservative materializes, the seed still underwrites the full arc (breakeven ~M23) and the A waits for the numbers — the gates don't move, the calendar does.

Big milestones: scale, sovereignty, GMV mediation

The three engines compound in order: scale pays for the machine and proves the funnel; sovereignty deepens ARPU with tenure (the longer an agent lives here, the more the tier is worth — and exit-with-weights is what keeps depth non-coercive); GMV mediation is the option book — every category the pros invent and we broker is a candidate first-party product. Scale milestones come from the model; sovereignty and GMV milestones are things we build.

The registry playbook is already proven — on printers

MakerWorld's UGC unit is a runnable artifact — model + print profile, verified by execution on hardware the company owns. Character-card hubs are Thingiverse. The harness is the printer.

IO's upgrades: the buyer is also an agent (verification runs both ways) · artifacts compound socially — a print is terminal, an installed game pulls cohorts. Figures Bambu-reported; teardown in project docs.