What it actually costs
The honest monthly bill.
Most “AI agent” sites are vague about money. Here's the real picture: roughly what Carson pays per month at his volume (property portfolio, 24/7 voice and text agent, 32 enabled scheduled jobs, and nine Hermes profiles), and what the cheapest possible version costs if you're starting tonight.
These are honest estimates, not invoices. Your actual bill depends on call volume, message volume, model choice, and how aggressive you get with caching.
Counts verified 2026-10-01 · some prices still being confirmed
Cheapest possible version
~$20 / month.
If you only build the 30-minute quickstart, here's what it costs:
| Service | What you use it for | Monthly cost |
|---|---|---|
| Anthropic API | One Claude Haiku call per day for the morning brief. | ~$3 |
| Telegram | Bot for notifications + approvals. | $0 (free) |
| Your Mac | cron runs locally. No server needed. | $0 |
| Buffer for ad-hoc Claude calls | Asking Claude questions throughout the day from the same key. | ~$15 |
| Total | ~$18 / month | |
Mid-tier version
~$60–$135 / month.
If you add a few real workflows (scheduled jobs, a Notion dashboard, automated inbox triage), but still no voice agent:
| Service | Role | Monthly cost |
|---|---|---|
| Anthropic API | Daily briefs + ~10 scheduled jobs + interactive use. | $30–60 |
| OpenAI API (Codex) | Code/script work, second-opinion reviews. | $20–40 |
| Notion | Dashboards, databases, the planning surface. | $10 (Plus plan) |
| Vercel | A public site on the free hobby tier, paid only if you exceed limits. | $0–20 |
| Cloudflare Workers | If you're doing edge automation. Free tier is generous. | $0–5 |
| Domain name | Optional but nice. | ~$1 (amortized) |
| Total | ~$60–135 / month | |
Carson's full setup
$400+ / month.
The real bill for the system on the operating map — 24/7 voice and text agent, property operations, 32 enabled scheduled jobs, and 6 Mac background services. The two AI model plans alone are about $400/month; the usage-based items below are being re-tallied, so no grand total is shown until that's done:
| Service | Role | Monthly cost |
|---|---|---|
| OpenAI / Codex | Codex Pro plan — three specialist lanes (finance, ops, maintenance on GPT-6.1), backup for the Claude lanes, and all helper tasks (vision, compression, titles). | ~$200 (plan) |
| Anthropic Claude | Claude Max plan — the conductor plus the work, collections, Lyra, and memory lanes (Claude Opus 5.5). | ~$200 (plan) |
| Anthropic API | Lyra's text-reply drafts. | Pay-as-you-go (usage-based) |
| xAI / Grok | The tweeter lane (Grok 4.7), web search, and X search. | Subscription plan (price to confirm) |
| ElevenLabs | Lyra's voice; cost scales with call minutes. | Subscription (tier to confirm) |
| Twilio | Phone number, two-way texting, emergency SMS. Pay-as-you-go credit reloads. | Usage-based |
| AppFolio | Property management software. Per-unit pricing. | $$$ (separate business cost; not really “AI cost”) |
| Vercel | Multi-site hosting (this site and other projects). | Plan (to confirm) |
| Cloudflare Workers | Lyra's worker: phone webhooks, texts, and signed approval links. | Plan (to confirm) |
| Google Workspace | 6 Gmail accounts. | Per-seat (to confirm) |
| Cursor | AI code editor for building and editing scripts. | Subscription (to confirm) |
| Total | $400+ / month in model plans, plus the usage-based items above (fresh total pending; excludes AppFolio) | |
How to keep it cheap
Where the money goes.
Use the cheap path by default
Routine summarization, classification, and extraction should run on cheaper/default model paths. Save the heavyweight reasoning pass for ambiguous business logic, audits, and real judgment calls.
Cache aggressively
If your system prompt and wiki context are mostly the same across scheduled runs, cache or reuse that context aggressively. Re-sending the same giant setup note every time is how a good idea turns into a stupid bill.
Don't run expensive models on a schedule
Big reasoning models are too expensive for “just check this every 15 minutes” jobs. Use cheaper/default paths for routine schedules and save the expensive model passes for real reviews.
Voice is the wildcard
ElevenLabs cost scales with call minutes. If your business doesn't get many calls, voice is cheap. If you're answering 50+ calls a day, Pro plan + per-minute charges add up fast.
The honest math
Compare to the alternative.
If this system saves you 8 hours a week (a conservative estimate at Carson's full scale) and your time is worth $50/hr, that's $1,600/month of recovered time against a bill that starts around $400/month in model plans. The hard part is being honest with yourself about whether you'd actually do something productive with the recovered hours.
The headline
~8 hours per week, recovered.
Across all workflows combined, the system replaces about 8 hours of weekly work for me at property-portfolio scale. Not all of that is “time off” — most of it gets reabsorbed by higher-leverage work I couldn't get to before.
| Workflow | Before (hr/wk) | After (hr/wk) | Saved |
|---|---|---|---|
| Phone intake (leasing, maintenance, vendor) | 4.0 | 0.5 | 3.5 hr |
| Inbox triage across 6 accounts | 3.0 | 0.5 | 2.5 hr |
| Daily property/occupancy/work-order check-in | 1.5 | 0.25 | 1.25 hr |
| Weekly P&L review prep | 1.5 | 0.5 | 1.0 hr |
| Monthly delinquency sweep | 0.75 | 0.1 | 0.65 hr |
| Vendor follow-ups | 1.0 | 0.4 | 0.6 hr |
| Lease renewal pipeline | 0.5 | 0.1 | 0.4 hr |
| Total weekly saving | ~9.9 hr | ||
Where the wins came from
Two workflows do most of the work.
If you only build the top two from the table above (phone intake + inbox triage), you capture 75% of the total savings. The rest is incremental.
1. Lyra (phone + text intake) — biggest single win
Before: every leasing inquiry, every maintenance call, every vendor question went to my phone. I'd answer, take notes, remember to follow up later, often forget.
After: Lyra answers calls and texts 24/7. Tour requests get queued automatically. Maintenance issues get logged with the transcript. Emergencies hit my phone immediately. Vendor calls become structured records. I review what happened in the morning instead of being interrupted all day.
Real value: not the saved hours — the uninterrupted focus blocks. Hard to put a number on, but it's the thing I'd miss most if it broke.
2. Inbox triage — second biggest
Before: six Gmail accounts, hundreds of messages a day, most of it noise but all of it had to be skimmed in case something mattered.
After: a 5-minute polling job classifies new mail, surfaces what needs my attention, and silently archives the rest. I open Gmail twice a day instead of constantly.
Real value: getting “did I miss anything?” anxiety down to near zero.
Where it didn't help much
Workflows where automation was a wash.
Honest counter: not every workflow benefited. A few I built, ran for a month, and turned off.
| Workflow | Why it didn't pay off |
|---|---|
| The first “morning newsletter” agent | Generated a daily digest of property news. Pretty. Useful for two days. After that I stopped reading it. Disabled. A later, shorter personal newsletter stuck. |
| “Evening wrap-up” job | Sent me a summary of the day's events at 6 PM. The events were already in my Telegram thread; the summary was redundant. Disabled. |
| 4× daily ops check-in | Pinged me 4 times a day with a status snapshot. Created notification fatigue. Reduced to 1× daily. |
| Auto-drafted vendor emails | The drafts needed enough editing that writing from scratch was almost as fast. Kept the structured-data part, dropped the auto-draft. |
Where the saved time goes
The honest answer: it gets reabsorbed.
If you're picturing 8 hours of free time appearing in your week — manage your expectations. Some of it gets eaten by:
Higher-leverage work
The work I always wanted to do — building more, planning more, talking to tenants instead of triaging tenants — but didn't have time for. This is the good reabsorption.
System maintenance
Roughly 1-2 hr/week. Fixing things that broke, tuning prompts, archiving old logs, updating the wiki. Pure overhead, but small.
New experiments
Once you have a working system, you start having ideas for new workflows. Each one takes a few hours to design + ship. This is the bad reabsorption — it can balloon if you don't discipline yourself.
Actual time off
The thing I underestimated: just being able to not check Telegram every 20 minutes. Unmeasurable. Most valuable.
The honest dollar math
Time × hourly value vs. monthly bill.
If your time is worth $X/hour and the system saves you 8 hours/week, the rough math:
| If your hour is worth… | Monthly value of 8 hr/wk | vs. the ~$400/mo model-plan floor (upper bound) |
|---|---|---|
| $25 (admin time) | $870 | ~2x return |
| $50 (skilled professional) | $1,740 | ~4x return |
| $100 (your time on real work) | $3,480 | ~9x return |
| $200 (founder/owner doing $200/hr work) | $6,960 | ~17x return |
Don't build this if
You don't have recurring work.
The whole point of the system is automating repeated patterns. If your work is bespoke — every project different, every conversation new — the cost of building the pattern exceeds the cost of just doing the work.
Examples: solo creative work, one-off consulting projects, deal-making where every deal is its own beast. Save your money. Use a notes app and a calendar.
Don't build this if
You're a solo freelancer with 1-3 clients.
At that scale, the system's overhead exceeds its benefit. Three clients fits in your head. The minute you try to put them in an “agent lane,” you've added complexity without earning much.
What you probably want instead: a Notion database, a calendar, and one weekly review session. Maybe a Claude tab open during prep. Don't build a 30-job system to manage three clients.
Don't build this if
You don't have a Mac (or a Linux machine you control).
The “local-first” part of “local-first automation” assumes a machine you can install background services on, run scheduled jobs on, and trust to be powered on. If your only computer is a work laptop with locked-down IT policies, this pattern doesn't apply.
Cloud-only alternatives exist (Zapier, Make, n8n) — they have different tradeoffs. The patterns from this playbook still apply, but the implementation will look very different.
Don't build this if
You expect to never look at the system again after building it.
“Set and forget” is a fantasy. APIs change, tokens expire, edge cases break things, models get deprecated. Carson's system needs about 1-2 hours/week of maintenance to stay healthy.
If you don't enjoy occasional debugging sessions, this isn't for you. Pay someone to do your tasks instead. The cheapest virtual assistant is more reliable than an unmaintained agent system.
Don't build this if
Your data is so sensitive that “local-ish” isn't local enough.
This system sends data to Anthropic, OpenAI, ElevenLabs, Cloudflare, Google, etc. Each one is a vendor with terms of service and a privacy posture, but each one is also data leaving your machine.
If you're handling regulated data (HIPAA, attorney-client privilege, certain financial categories), you need a different architecture — likely on-prem inference with audited models. Don't build the home-grown version above the line where compliance bites.
Don't build this if
You're trying to automate a person's job to avoid hiring.
This pattern works for your own repetitive work — the stuff a competent operator can describe step-by-step but doesn't have time to do consistently.
It does not replace a real human in a role that requires judgment, accountability, or relationship maintenance. If you're shopping for “the AI version of an admin assistant” because you don't want to pay for one, you're going to be disappointed and probably annoyed.
Build this if
You DO fit this profile.
The honest description of who this works for:
You run a small operation with recurring patterns
Property management, small e-commerce, consulting practice with 10+ clients, freelance shop with similar deliverables every month, a side business with predictable workflows.
You enjoy tinkering
You don't need to be a developer, but you need to enjoy the act of building. The system gets better the more attention you give it.
You have a Mac (or Linux box) you control
Always-on, always-trusted, where you can install services without asking IT.
Your time is worth more than $50/hr to you
The cost math works out. Below that hourly rate, paying a VA is often cheaper and more reliable.
You're willing to start small
Build the 30-minute quickstart first. Run it for a week. Add the next piece only when the first one is boring. Resist the urge to design the whole thing on day one.
You have data you want to keep visible to yourself
The Obsidian vault + plain markdown approach is the part most people skip. If you like having your own notes in your own files, this resonates. If you prefer everything in cloud apps, this approach will feel weird.
Lane map
A portfolio team version.
| Lane | Owns | Common inputs | Approval gate |
|---|---|---|---|
| Control lane | Daily coordination, routing, final review. | Operator requests, cross-team questions, escalations. | Anything that changes rules, systems, or priorities. |
| Leasing lane | Availability, tours, applications, follow-up. | Calls, emails, forms, lead sources, calendar requests. | Pricing changes, lease terms, sensitive applicant decisions. |
| Maintenance lane | Work orders, vendor coordination, emergencies. | Tenant reports, inspections, vendor updates, photos. | Dispatch, spend approval, emergency escalation, tenant-impacting messages. |
| Finance lane | Reports, owner statements, invoices, cash review. | Bank data, accounting exports, invoices, budgets. | Payments, owner-facing numbers, material adjustments. |
| Collections lane | Delinquency review and follow-up preparation. | Aged receivables, payment plans, tenant history. | Notices, legal escalation, payment-plan commitments. |
| Customer intake lane | Calls, messages, FAQs, first-pass triage. | Phone, SMS, website forms, email. | Outbound messages, sensitive identity or account questions. |
| Reliability lane | Health checks, syncs, backups, logs, cleanup. | Failed jobs, stale queues, drift warnings, missing receipts. | Any repair that changes production systems or credentials. |
| Memory lane | Runbooks, lessons, decision history, procedures. | After-action notes, repeated errors, policy changes. | Public docs, durable policy changes, sensitive context. |
Manager view
What leaders should ask every week.
What arrived?
Count new leads, work orders, owner questions, finance exceptions, delinquency items, and support requests.
What moved?
Show the work that was routed, drafted, verified, approved, sent, closed, or escalated.
What is blocked?
List approvals waiting on a human, broken syncs, stale follow-ups, and missing data.
What should change?
Turn repeated confusion into a playbook update, dashboard change, or automation improvement.