Cheapest model that passes
Eval gate · not a beauty contest
Python that stays up for your agents.
Persistent Python for AI agents.
Spin up a fixed-rate instance. Your agents install packages, run for days, spawn more agents — model-agnostic. You don't manage servers. You buy compute that stays warm.
From one-off tool calls to persistent computational workflows.
instance-medium
Persistent interpreter · fixed $/hr
Uptime
Tool calls landing
- 12:04:21
pandas.readallow - 12:04:19
retrieveallow - 12:04:18
python.execallow - 12:04:31
net.rawdeny
Warm cache
numpy · pandas · httpx
Rate
US$0.08 / hr
Isolation
System container
FIXED RATE
$0.08 / hr
PACKAGES
Still installed
DESTINATION
python.exec
Illustrative session
3 asks · 22 agent steps
Duration
16 min
Assistant steps
22
Tool-call share
66.7%
Session
US$1.97
Pattern from a typical agent session dashboard. Digits are labelled illustrative — not a PySmith customer bill.
Where calls land
Pieces make each other better
Agents route tool calls through MCP. PySmith is the always-on Python desk those calls land on — packages warm, state intact, bill predictable.
Score live runs, not just test sets
Tracing + evals
Govern tool access in one place
MCP Gateway + agents
Agent
MCP
python.exec→PySmith runtime✓unsafe.shell✕In the field
Where a persistent Python desk earns its keep
These are evidence-led concepts, not deployments we claim as our own. Each story borrows a published operating problem, then shows where PySmith sits: the always-on Python desk behind MCP tool calls — never the system that signs the decision.
Evidence-led conceptDecision support
Decision-grade finance
Keep evidence, computation and review history together long enough to inspect, challenge and reproduce.
5.7% → 3.7%
>98%
>5,000
Evidence-led conceptMulti-agent operations
Shift-scale mining
A persistent, inspectable Python environment for testing coordinated decisions across a full operational horizon.
+5.56%
Mining-Gym
No published %
Evidence-led conceptPlanning continuity
Living supply plans
Keep the planning computation available through every replan — people and solvers keep the decision.
>98%
~40%
~93%
Agents route tool calls through MCP. PySmith is the always-on Python desk those calls land on — packages warm, state intact, bill predictable.
The problem
Agents do not sleep like web apps
01
Ephemeral sandboxes forget
The room is reset. Packages vanish. Week-long agents do not enjoy unpacking every morning.
02
A raw VM is a second job
You wanted Python. You got patching, SSH, and a disk that fills up on a Saturday.
03
Per-second bills surprise you
Honest when idle. Noisy when a child process runs for days. Spreadsheets prefer a fixed rate.
04
Agents need a socket, not a ticket
MCP is how they find tools. Exec Python should not require a human with a console.
How it works
A desk. A rate. A socket.
Controlled Python sandboxes for multi-step AI agent work.
CONNECTINSTALLSTAY UPSPAWNPYTHON
01
Connect
Point your agent at the PySmith MCP connector. No cluster YAML.
02
Install
uv or Pixi on a warm cache. Common libraries are already close.
03
Stay up
The interpreter, the files, the child processes — still there tomorrow.
04
Spawn
Agents exec Python and start more agents. Model-agnostic. You buy hours, not drama.
What you buy
Persistent Python. No infra.
Managed Python sandboxes for stateful AI agent workflows.
Days and weeks, not session caps
Built for agents that keep a workspace. Continuous processes, not a 24-hour hotel room.
A number you can budget
Small, Medium, Large. The clock is the product. Idle-heavy work should use someone else.
Agents exec Python
The connector is the API. Your planner does not SSH. Your runtime does not care which model called it.
uv / Pixi, not a blank pip
Install once. Come back. We cache per tenant. We do not silently share wheels across customers.
You do not run the host
No EC2 console. No security group folklore. You buy a desk. We keep it standing.
Not Firecracker. Not 100k sandboxes.
System containers, fixed SKUs, Python first. Modal, E2B and Northflank already exist. We sell the leftover wedge.
The wedge
Always-on, fixed-price, MCP-first Python
Modal, E2B and Northflank already cover burst scale, Firecracker sandboxes, and long-running volumes. We do not pretend otherwise.
The leftover job is smaller: indie builders who want a warm Python desk for days and weeks, a known hourly rate, and an MCP socket — without babysitting EC2. Thin wedge. Useful desk.
Read the product storyServerless sandbox → hours
Raw VM → ops
PySmith → desk
Raw VM → ops
PySmith → desk
Isolation: system containers, not microVMs.
Scale: a desk, not a fleet of 100k.
Scale: a desk, not a fleet of 100k.
Pricing sketch
Fixed SKUs. Illustrative until launch.
Small
USD 0.04 / hr
A quiet desk for one long-running agent. About USD 29 / mo if you leave it up.
Medium
USD 0.08 / hr
The default. Packages stay put. Agents spawn children. About USD 58 / mo if you leave it up.
Large
USD 0.14 / hr
Heavier jobs that still need a known monthly number. About USD 102 / mo if you leave it up.
Illustrative launch pricing, not a quote. Rates will move. No fake discounts, no invented free-tier minutes.
Get a desk
Python that stays up for your agents.
Spin a fixed-rate instance. Install what you need. Leave it running. Your agents already know how to use a tool socket.
Recent Blog Posts
View all →Why agent builders keep overpaying for servers
Teams buy a full VM so Python packages survive overnight. Most of the bill is idle time. The data-platform world already has a name for this leak.
Your agent needs a desk, not a data centre
Ephemeral sandboxes forget installed packages. Raw VMs want a human on call. Long-horizon agents need a warm Python desk that stays set up.