For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The first economic fact about Markov Blanket is that email is the cheap part.
The public pipe is already a commodity. Cloudflare's current email stack can route inbound mail freely, send the first 3,000 outbound messages inside a $5/month Workers Paid floor, and then send more at $0.35 per thousand. Resend, Mailgun, AgentMail, and SES price the same layer differently, but they do not change the shape of the business. They sell addressability, delivery, logs, webhooks, and reputation.
Markov Blanket sells the boundary.
The bill begins when a capable model decides what should cross that boundary, drafts an outbound artifact, chats with the user, remembers a correction, or compacts a growing graph. It gets worse when the product uses SMS, because Twilio charges for every segment in both directions. It becomes absurd when internal creature-to-creature traffic is treated as public email instead of a private webhook.
That is the unit-economics spine: rent the pipe, own the membrane, meter the creature's outward force.
The useful comparison asks what each provider is actually selling.
| Provider | Useful price point | What it buys |
|---|---|---|
| Cloudflare Email Service | $5/month Workers Paid; 3,000 outbound/month included; then $0.35/1,000 | The best first pipe because the current primitive bundle already wraps Cloudflare into narrow agent-native verbs. |
| AgentMail | $20/month for 10 inboxes and 10,000 emails; $200/month for 150 inboxes and 150,000 emails | Agent-native inboxes, threading, webhooks, search, and MCP surface. Worth buying if it saves build time. |
| Resend | $20/month for 50,000 emails; overage $0.90/1,000 | Pleasant developer-first sending and receiving. Good fallback if Cloudflare's mail surface is too young. |
| Mailgun | $15/month for 10,000; $35/month for 50,000; overage from $1.80 or $1.30/1,000 | Deliverability operations, logs, and an email-company-shaped surface. |
| Amazon SES | $0.10/1,000 outbound | Cheapest commodity sending, with more AWS and deliverability surface to hold. |
| Self-hosted SMTP | $5-$20/month for a small server before labor | Cheap in server dollars; expensive in reputation, abuse, monitoring, and deliverability. |
At 9,000 outbound emails per month, Cloudflare is about $7.10 including the $5 floor. Resend is $20. Mailgun Basic is $15. SES is $0.90 before the AWS surface. AgentMail is $20 if the account fits inside its Developer tier.
At 90,000 outbound emails per month, Cloudflare is about $35.45. Resend sits roughly in the $56-$90 band depending on plan choice. Mailgun sits near the $87-$90 band. SES is $9. AgentMail moves to the $200 Startup tier.
Those numbers make self-hosted SMTP look tempting only if the table stops at dollars. It should not. Gmail and Yahoo require SPF, DKIM, DMARC, reverse DNS, low complaint rates, and bulk-sender hygiene. A young product sending personal creature mail does not want its first scarce asset to be IP reputation. Public email deliverability is already a business. Markov Blanket should not become that business before it has proven the creature.
The right in-house move is the internal message bus. Between Markov Blanket systems, use webhooks and POST requests. Keep email semantics at the human boundary, where the world already speaks email. Do not make internal thought pay MX-record tax.
HEY is the closest product precedent because it understood that an inbox is a permission surface before it is a storage surface.
HEY charges $99/year for personal email. HEY for Domains charges $12/user/month, with the first user at $10/month. Its important move is the Screener: each sender is admitted or refused, then routed into Imbox, The Feed, or Paper Trail. The user is paying for a boundary she can feel.
That is the posture to copy, not the closure. HEY keeps the opinion inside HEY. Markov Blanket should let the user's model leave. The sender decision should update readable filter, voice, memory, and constitution files the user can inspect, argue with, export, and carry forward.
HEY also sets a floor. A serious email product can charge about $8.25/month for personal email or $10-$12/month for custom-domain email without an AI creature. Markov Blanket is asking to run a model of the user's boundary. It should not pretend that is a $5 utility.
The monthly cost per active user is:
cost =
fixed infrastructure / active users
+ mail transport
+ SMS transport
+ storage and indexing
+ model triage
+ model chat
+ model outbound drafting
+ model compaction
+ support and refund reserve
Gross margin is:
gross_margin = (price - cost) / price
For 50% gross margin, a $20 plan can spend only $10/user/month. A $75 plan can spend $37.50/user/month.
That threshold decides the product.
These sizing constants expose the slope. Final architecture can move them; the direction is what matters.
| Work unit | Token assumption | Model route | Unit cost |
|---|---|---|---|
| Inbound triage signal | 1,500 input / 300 output | Claude Haiku | about $0.0030 |
| App chat turn | 4,000 input / 700 output | Claude Sonnet | about $0.0225 |
| App chat turn, cheap route | 4,000 input / 700 output | Claude Haiku | about $0.0075 |
| Outbound draft | 2,000 input / 500 output | Claude Haiku | about $0.0045 |
| Outbound draft, stronger route | 2,000 input / 500 output | Claude Sonnet | about $0.0135 |
| 3,000-node compaction | 3M input / 100k output | Claude Haiku | about $3.50 |
| 3,000-node compaction, stronger route | 3M input / 100k output | Claude Sonnet | about $10.50 |
Run those constants through the first usage bands:
| Scenario | Monthly behavior | Estimated monthly cost | $20 margin | $75 margin |
|---|---|---|---|---|
| First-email prototype | 30 inbound, 30 outbound, no chat | $5.22 | 74% | 93% |
| Low-touch paid trial | 900 inbound, 500 chat turns, 300 outbound | $20.30 | -2% | 73% |
| Intended average with webhooks | 3,000 inbound, 500 chat turns, 900 outbound, 1 compaction | $32.80 | -64% | 56% |
| Power email user | 3,000 inbound, 500 chat turns, 9,000 outbound, 3 compactions | $78.35 | impossible | -4% |
| 3,000/day outbound limit | 3,000 inbound, 500 chat turns, 90,000 outbound, 30 compactions | $565.70 | impossible | impossible |
The first email pays for itself. A $20 first-user month can cover Cloudflare and a small amount of model work. If the broader Hari account is already paying the Cloudflare floor, the marginal first-email cost is even closer to zero.
The $20 plan does not survive as an uncapped lived creature. It can be a trial, a door, or a bounded basic tier. It cannot include 100 inbound signals/day, 500 app turns, real outbound drafting, compaction, support, and a 50% margin.
The $75 plan works for the intended early average if external outbound action is held near 30/day, internal systems use webhooks, and model routing is disciplined. It leaves roughly 56% gross margin under the sizing assumptions above. That is real, with little slack.
The $75 plan breaks for a power user sending 300 model-drafted external emails/day. That usage needs a higher tier, a meter, cheaper routing, or a product rule that most outbound action stays internal until it must cross into the world.
The 3,000/day outbound limit belongs to enterprise automation.
Twilio US long-code SMS is $0.0083 per inbound or outbound segment. A 6,000-message one-way month is about $49.80 before model cost. If a "turn" means one inbound and one outbound segment, 6,000 turns are about $99.60 before model cost, number rental, registration, media, carrier fees, and long-message segmentation.
SMS cannot be silently bundled into a $75 plan at power-user volume. It should be metered, an add-on, or reserved for high-value lifecycle moments. The product can subsidize a few onboarding or rescue texts. It cannot subsidize 6,000 ordinary text turns without becoming a gym subscription with the wrong power-user math.
The 3,000-node graph threshold sounds large because it is cognitively large. As storage, it is tiny.
Cloudflare D1's paid tier includes 50 million row writes/month and 5GB storage before overage. R2 includes 10GB-month free and then charges about a cent and a half per GB-month. Embedding 3,000 nodes of 1,000 tokens each with a cheap embedding model is measured in cents.
The expensive part is asking a model to understand the nodes.
That means the memory policy should meter compaction passes, retrieval depth, and graph rewriting, not raw bytes. Store generously. Compact deliberately. Spend frontier cognition where the boundary is uncertain.
At the intended $75 average:
price = $75
estimated cost = $32.80
gross profit = $42.20/month
LTV at 10% monthly churn = $422
LTV at 50% monthly churn = $84
At a cheap-routed $20 trial with about $12.80 cost:
gross profit = $7.20/month
LTV at 10% monthly churn = $72
LTV at 50% monthly churn = $14.40
Churn does not change gross margin. It changes whether gross margin can carry acquisition, support, refunds, and mistakes. If customer acquisition cost is near zero, $20 can be a learning price. If paid acquisition appears, $20 has almost no room for it. CAC should be treated as a gift, not a law of nature.
Use Cloudflare first. The current primitive bundle already chose that shape: narrow agent-native mail verbs over Cloudflare Email Sending and Email Workers. That fits the agent-native-tooling thesis. Raw SMTP is too broad a surface for the agent. Cloudflare gives the creature an address without making the team an email company.
Keep Resend, Mailgun, SES, and AgentMail as live options. Resend is the pleasant developer path. Mailgun is the deliverability-ops path. SES is the cheapest commodity path. AgentMail is the agent-inbox acceleration path if its inbox, thread, search, webhook, and MCP primitives save enough build time to justify the higher tier.
Bring in-house the membrane: alias purposes, sender gates, filter files, voice correction, relationship memory, graph compaction policy, model routing, public-safe ledgers, and internal webhooks. Those are the product. Those are the parts that compound.
Do not bring in-house public SMTP for v1. Build private message-passing infrastructure instead. Public deliverability is a reputation business. Markov Blanket is a boundary-and-memory business.
Subsidize the first email, onboarding, light inbound triage, and the correction loop. Those teach the creature and prove trust. Eat occasional compaction and storage. Eat refunds when the failure report is specific enough to become product work.
Meter outbound automation, SMS, deep model turns, high-frequency compaction, attachments, web search, tool calls, and any external action that creates obligation. Those are where power users become real cost centers.
Price the lived creature at $75 or above. Let $20 be a serious trial, not the actual animal. The $20 plan can open the door. The $75 plan can keep the creature alive if the architecture refuses to turn every internal thought into an email and every user whim into a Sonnet call.
The economics do not say "be cheap." They say where generosity belongs. Give away the pipe-level affordance because the pipe is cheap. Protect the model-mediated boundary because that is where care and cost are the same object.