Ways to run an AI agent

How Plori compares.

Plori runs AI agents in cloud Workspaces. Give an agent a task and keep its files for the next run. Use it to hand work to an agent Plori operates, from the web app or from your own agent over MCP. Use a sandbox API such as E2B to run code for an agent you build, a VPS to operate the agent yourself, and a local agent when the work needs your device. If your current cloud agent already handles the task, keep it.

Sign in with Google to start.

Who runs the agent, and what do you manage?

§01
OptionWho builds and operates itWhat persistsCosts to includeResults, scheduling, and approvals
PloriPlori operates the agent. Use the web app, or let your own agent (Claude Code, Cursor, Codex) drive Plori's agents over MCP.Only Workspace files persist, with saved revisions, the conversation, and run history. Software installed outside the Workspace and running processes do not carry over to the next run.Credits, a prepaid balance in US dollars, pay for model usage and run time ($0.01 per minute, billed by the second). Each run gets its own VM; in our measurement, VMs stopped a median 1.739 s after run end and idle windows incurred $0 in charges. Include subscriptions and extra disk. Idle billing measurements.Replies in the conversation, files in the Workspace. Several workers can run in one Workspace at once, each on its own copy of the files. Plori merges changes and handles conflicts. Scheduled runs and agent-built workflows. Irreversible actions outside the agent's environment wait for the owner's approval in the web app.
Sandbox APIs (E2B, Daytona, Modal)The provider operates code execution. You build the agent and the application around it.Depends on the vendor and mode. E2B pause can keep files and memory. Daytona container stop keeps files, not memory. Modal snapshots keep files for a later sandbox.Compute, storage, any plan fee, and your model and tool providers.Your application connects execution output to users, schedules, credentials, and approvals.
A VPS you manageYou install and operate your own agent on a rented server.Disk follows the server and its backups.Server rental, storage, models, tools, and your time. Check whether a stopped server still bills.You build result delivery, scheduling, permissions, updates, and monitoring.
A local agent on your laptopYour agent process runs on your own device.Files stay on the device. Sleep, shutdown, or a process failure can interrupt a run.Hardware and electricity, plus model and tool charges.Unattended work depends on the device staying available.
Cloud agents in another product (for example, Cursor)The vendor operates the agent, with its own setup and integrations.Cursor cloud agents run in isolated cloud VMs. Automations keep notes in MEMORIES.md across runs.Cursor bills Automations as cloud agent usage. Check plan fees and model usage.MCP support, PRs with artifacts, and Automations on cron schedules or events from GitHub, Slack, Linear, and webhooks.

Plori supplies the agent. A shell, network access, persistent files, memory, and scheduled runs are common in this category. With Plori, your own agent can drive Plori's agents over MCP, including several workers in one Workspace, and Plori's agent builds its own workflows in web chat.

Which setup fits your task?

§02

Building an agent into your own application?

Use a sandbox API such as E2B, Daytona, or Modal for isolated code execution in an agent you build, and connect its results to users yourself. Use a VPS when you need control of the server and can maintain it. Plori supplies its own agent and does not deploy your agent framework.

Handing off a test run or research task?

Try the cloud agent in the product you already use first. Your current agent can also hand the work to Plori's agents over MCP. It can split the work across several workers and collect the result later, and the Workspace keeps the files. Work that needs local-only data or a desktop application stays on your device. A Plori run waits for the account owner's approval before an irreversible action such as a git push. Each plan limits how long one run can last.

Need a report on a schedule?

A Plori scheduled run starts one agent task later. For a repeating routine, describe it in web chat and the agent builds a workflow. After you activate it, the workflow runs on a schedule or a webhook with no agent turn.

What did we measure on Plori?

§03

The wake-latency and worker measurements cover Plori production release v350 on 2026-10-14 and 2026-10-15, with run time billed by the second. The other measurements cover release v349 on 2026-10-14, under started-minute compute billing. We did not measure the other products in the table.

  • On v350, for a one-word reply, median time from message to first reply text was 2,538 ms with a VM from the warm pool (n = 120, serial runs). It was 3,509 ms with a new VM restored from a template (n = 29). Each of those 29 runs started together with another run while the pool kept one ready VM. Wake latency.
  • VMs stopped a median 1.739 s after run end (n = 18), and 366 checks over six 30-minute idle windows found no agent VM and $0 in idle charges. Idle billing.
  • Median time from run end to saved revision ready was 1,769 ms for 1 MiB, 4,076 ms for 64 MiB, 8,434 ms for 256 MiB, and 23,042 ms for 1,024 MiB (n = 5 per size). Workspace save time.
  • On v350, dispatch to all worker changes accepted took 9.9–17.9 s with one worker and 20.8–30.9 s with four workers in one Workspace (three repeats each). Accepts ran serially at about 1 s each. Worker timing.

Questions, answered.

§04

Should I use Plori or a sandbox API like E2B?

Use a sandbox API when your application needs code execution and you build the agent around it. Use Plori when you want to hand tasks to an agent Plori operates, from the web app or from your own agent over MCP.

When is Plori not the right choice?

Use another setup when you need to deploy your own agent framework, require on-premises execution, or need a graphical desktop. Keep your existing cloud agent if it already handles the task.

Is persistence unique to Plori?

No. E2B, Daytona, and Modal each keep files in some lifecycle modes, and E2B pause and Daytona VM pause can also keep memory. Plori keeps Workspace files with saved revisions, but not software installed outside the Workspace or running processes.

What should I compare in the total bill?

Include the subscription or server fee, model usage, compute, paid tools, workflow runs, and storage. On every Plori plan, run time costs $0.01 per minute, billed by the second and paid from the same prepaid dollar balance as model usage. Each run gets its own VM; in our measurement, VMs stopped a median 1.739 s after run end and idle windows incurred $0 in charges. Your plan includes some disk, and extra disk costs money. With your own OpenRouter or provider key, the provider bills model usage and other Plori charges still apply.

Can I move my work to another setup?

You can download files or ask the agent to push work to your own git remote or storage. Moving files does not move the agent, running processes, saved memory notes, or connected-account credentials.