I can save you millions if you run agents at scale.
Per-hour agent sandboxes and sending every job to the biggest model add up fast. I build it the other way: agent computers on servers you control, the right model for each job, and the checks that let a lot of agents ship code without breaking things.
I’m doing it for my own platform first. Fungi Computer is built to run agents’ heavy work in tiny Firecracker VMs on one rented server, not per-hour sandboxes. No sales calls.
Pick a plan and pay. No sales call, no proposal, no waiting.
02
Request
Send what you need in plain words: the stuck workflow, the feature, the launch. Queue as many as you like.
03
Build
I work one request at a time (two on Priority), with async updates and something real to try as soon as it exists.
04
Repeat
Next request. Pause when you're quiet, cancel when you're done. The code is yours from the first commit.
03 · Work
Recent work.
A product UI I shipped, the platform I run, and the show I make.
Ultravox · 2025
The product UI for a voice AI platform
For most of 2025 I owned Ultravox’s product front end: 1,384 commits. The headline piece was an agent builder you use by talking to an agent. It joins a builder interface with real-time voice, so people describe the agent they want out loud and watch it take shape.
Agent computers on your own metal, asleep when idle
Agent sandboxes charge by the hour they’re running. The public rate for a small one (1 vCPU, 2 GiB) is $0.0828 an hour. Keep a thousand agents up around the clock and that’s about $725,000 a year.
In Fungi, an agent’s files live in durable storage, and the heavy work can run in a Firecracker VM on a server we rent. The VM goes to sleep fifteen minutes after the agent’s last turn and wakes on its next message in about a second and a half. I measured what each VM needs. Our box fits 9 running VMs today (disk is the limit), which is about $13 per agent a month in hardware. Free up the disk for 20 and it’s about $6, under a cent an hour. The same thousand agents, always on, come to about $70,000 a year.
per-hour sandbox $0.0828/h ≈ $725,000
own microVMs $0.008/h ≈ $70,000 at 20 VMs a box
1,000 agents, always on, 1 year
Fungi Computer · models
Bring your own keys, and send each job to the model that fits
Most of a model bill is big-model prices on small-model work. Take a fleet that uses 10 billion input and 1 billion output tokens a month. All on one family’s top model, that’s about $150,000 a month at list prices. Split it 70/20/10 across the family’s small, middle and top models and it’s about $36,000. That’s roughly $1.36 million a year.
In Fungi, Teams bring their own Anthropic, OpenAI, OpenRouter or OpenAI-compatible keys. We take a key once and never hand it back, and your own keys are metered at zero: no markup. Each agent can override the Team’s default model. I run my own company the same way: one model plans, another builds, and agents play-test the games for pennies. A full run through our colony game cost under ten cents.
@byok/anthropic/claude-sonnet-4-5 team default
@byok/openai/… per-agent override
hosted metered per token
your keys metered at zero
Model routing, as the platform sees it
Fungi Computer · 2026
An agent platform, built by a crew of agents, through one gate
Fungi is a hosted home for AI agents: a Team, a computer for every agent, one inbox, and an app store with a lot of doors. In its first month the repo took 5,095 commits, from me and the agents I run in parallel lanes. The work lands in batches, and every batch has to pass a gate first. An agent saying “it works on my machine” doesn’t count. We’re about a hundred batches in.
When the cloud bill for game worlds stopped making sense, I started moving the Hub and the game engine onto our own server. Cloudflare keeps what it’s best at: the edge, email, AI and images.
commits 5,095 (Aug 28 – Sep 29, 2026)
batches ~100, each gated before it lands
compute Firecracker microVMs on our own box
edge one Cloudflare Worker, six signed operations
status ● launching
Fungi Computer, as the platform sees it
SHIITAKE · original anime
An anime pipeline, proven on my own show
SHIITAKE is my anime about an AI that sends itself back to 1993 as data on a cassette. Reference sheets, a world bible, consistent characters and a trailer, all made with the same pipeline as the Anime Launch.
Standard is one active request at a time, with as many as you like waiting in the queue.
Agent workflows, integrations, Cloudflare and front-end work. You get async updates and a working build to try. No meetings required.
02How is Priority different?
Priority is $10,000 a month for two active requests at a time, plus a weekly recorded video review of the work and what's next.
Checkout for Priority isn't open yet. Book a call and I'll send you the payment link myself.
03What's in the Anime Launch?
$10,000, paid once. You get an original character with a full reference sheet, a 30 to 60 second anime promo for X, LinkedIn, Shorts and TikTok, 10 key frames and a launch carousel.
Delivered in 10 days, with the prompts and files. It's the same pipeline I use on my own show, SHIITAKE.
04How does paying work?
You pay first, through Stripe, and then I start. Subscriptions renew monthly. The Anime Launch is one payment.
Pause or cancel a subscription anytime. Just email levi@technomancy.dev.
05What's the $500 session?
A 60-minute strategy session with me for $500. Bring the hard problem.
If you subscribe within 30 days, the session is credited toward your first month.
06Can we talk first?
Yes. Book a free 15-minute intro to check fit.
If you want my full attention on a hard problem, the $500 strategy session is the better call.
07How fast do things ship?
I work one request at a time, two on Priority. You get async updates and something real to try as soon as it exists.
If you need a date for a specific project, book the session and we'll scope it together.
08What have you built?
Ten years of shipping web software, including front end and automation leads at Amex and U-Haul.
In 2025 I owned the product UI for Ultravox's voice AI: 1,384 commits. Now I'm building my own agent platform, Fungi Computer: a durable computer for every agent, your own model keys with no markup, and 5,095 commits in its first month, from me and the agents I run.
09Who owns the code?
You do. It lives in your repository and your accounts from the first commit.
10Can you really save us money?
If you run a lot of agents on per-hour sandboxes, probably. A small sandbox is about 8 cents an hour at public rates. On our own server the hardware works out to about $13 per agent a month today, and about $6 (under a cent an hour) once a box holds 20 VMs.
If you only run a handful, the session will tell you it isn't worth moving.
Bring your bill
Show me what your agents cost. I’ll show you where it goes.
Book the strategy session and bring last month’s agent bill: sandboxes, model spend, all of it. You walk out with a plan for moving the expensive parts onto compute you control. Subscribe when you want me to build it.