Cloud agents

A cloud agent is Lovely Agent, an AI agent app, running in a sandbox of its own. It chats with Claude, GPT or Gemini, runs commands, writes code and edits files, and uses Lovely Router's data tools and browser agent. Each agent keeps its files, sessions and settings between runs. You can have many agents, one per project or one per user of your own product, and reach each one in the browser or over the API.

1. 1

   Create an agent

   In Cloud agents select New agent, or call POST /v1/agents. Give it a name, and if it serves one of your users, your own id for them.

2. 2

   Use it

   Select Start, then Open: Lovely Agent opens at the agent's own address (agent.lovelyrouter.com/a/<agent id>/), already signed in. Or send it tasks with POST /v1/agents/{id}/runs. A stopped agent wakes up by itself when a run comes in or someone opens its page, unless you turn that off.

3. 3

   Stop it, or let it stop itself

   Select Stop when you are done. An agent also stops by itself after a period without use, and after the longest session you allow. Its workspace stays; the next start picks up where it left off.

What's inside#

• Lovely Agent with its web UI: chat, sessions, plan mode, subagents, skills, scheduled tasks, plugins and a terminal.
• Models: every chat model Lovely Router sells, already configured. The default is Claude Haiku 5.5, our least expensive; pick another per agent in the settings, or switch models inside the agent. You can also add your own model providers on its Models page.
• Tools: the data tools and the browser agent appear as MCP tools (mcp__tools__*), so the agent can search the web, get structured data from platforms such as YouTube, Reddit, Google Maps or Amazon in one call (it is told to prefer these tools to scraping the sites itself), make images or drive a browser. Add your own MCP servers in the settings.
• A Linux sandbox: Node.js 22, Python 3, git, build tools and ripgrep, with internet access. The agent has full access inside it; nothing outside it.

Pricing#

You pay for each agent's sandbox while it runs, plus what it uses at the usual prices.

The sandbox is charged every minute while an agent runs, from the moment it starts until it stops. Shared vCPUs run at full speed for bursts such as installs and builds, and at a lower baseline share during long CPU-heavy work; bursts cost nothing extra. Starting needs a balance that covers at least 15 minutes of the size (minimum $0.25). When your balance goes below zero your agents stop, and you get an email: unless you top up within 24 hours, their workspaces are deleted.

Between sessions an agent is kept ready, so its next start takes a few seconds and it picks up where it left off. Its disk starts at 1 GB and grows by 1 GB whenever it is 80% full (up to 20 GB), so a new agent costs about $0.66 per month while it is kept ready. An agent unused for 7 days is archived: only its files are billed, and its next start takes a few seconds longer.

Each agent shows what it has cost (sandbox time, and the models and tools its sessions used), and can have a budget: it stops, and won't start or run again, once it has spent that much.

White-labeled agents, which show your own branding to your users, cost 1.5× these sandbox rates (Small $0.075 per hour); models and tools keep their prices.

Settings#

Each agent has its own, in the console or with PATCH /v1/agents/{id}. They apply from the next start.

• Default model: the LLM new sessions start with. Default: Claude Haiku 5.5.
• Data tools and browser agent: on by default. Turn them off for an agent that only codes and chats.
• More tools (MCP servers): up to 8 remote MCP servers (streamable HTTP), each with a name, an https:// URL and an optional header such as Authorization: Bearer …. Header values are stored encrypted and never shown again.
• Instructions: up to 4,000 characters added to every session, for example your coding conventions.
• Size, Stop when idle for (5 minutes to 24 hours, default 30 minutes) and Longest session (1 to 24 hours, default 8 hours).
• Wake up when called (on by default): a stopped agent starts by itself when an API run comes in, when you open it, or when someone opens its page. Each wake-up is a normal session, billed as usual. A session used through the API stops about 5 seconds after its last run ends, whether a run woke it or it was already running; once someone uses the agent's page in a session, that session stops when idle instead (Stop when idle for, 30 minutes by default). To keep an agent running between API calls, send stop_when_done: false with the run. Turn the setting off to start the agent only yourself, from the console or with POST /v1/agents/{id}/start; your users' runs and page visits then find it stopped. This setting applies at once.

An agent counts as in use while someone uses its page, while it calls models or tools, while a run works, or while its CPU is busy (a build or a long command). An open tab alone doesn't count.

Agents for your own users#

Build agents into your product: one agent per user of yours, each with its own workspace, all paid from your balance.

1. Create an agent per user with your own id for them: {"user": "cust-42"}, plus any metadata. List a user's agents with GET /v1/agents?user=cust-42.
2. Give your user an agent token instead of your API key (POST /v1/agents/{id}/tokens). It reaches that one agent only, and expires by itself (1 minute to 30 days):
   • scope api: your user (or your backend acting for them) sends tasks with the agent's runs and reads its status. Use this when your users shouldn't see the web page;
   • scope ui: comes with a link that opens the agent's web page, signed in.

   With Wake up when called on, your user's run or visit starts the agent when it is stopped; they can't start it any other way.

3. Set a budget per agent to cap what each user can spend, and watch each agent's spend in the console or the API.
4. Show your user what the agent made: a run lists the files it delivered, and a share link opens a file or a whole web page without your user signing in.
5. Put your own product on the agent's page: give it one of your brandings, and your users see your name, logo and colors instead of ours.

Revoke an agent's tokens at any time; deleting the agent, or revoking the API key that created a token, ends it too.

Workspaces#

Each agent has its own workspace: its home folder with its projects, sessions, settings, skills and installed plugins, created at the first start and kept between runs. Sessions started over the API show up in its web page too, grouped as API Requests. Storage is billed daily: while the agent is kept ready, then, after 7 days without use, for what its archive holds (prices). Deleting an agent deletes its workspace for good and stops the storage charge. Workspaces are also deleted 24 hours after your balance goes below zero, unless you top up before then.

Files and shared pages#

What an agent makes (documents, reports, charts, data, a web page or a whole static site) stays in its home folder, and you can get at it three ways:

• In its web page: the files the agent hands over appear at the end of its reply. Select one to preview it; HTML pages run in a sandboxed frame.
• Over the API: a run lists the files it delivered (files, each with a download url), and GET /v1/agents/{id}/files/… lists or downloads anything in the home folder. That works while the agent runs (live) and while it is stopped (as saved at its last stop), without starting it.
• As a link anyone can open: Share in the console's file list, or POST /v1/agents/{id}/shares, makes a link that needs no sign-in. Share a file, or a folder to publish it as a website: the link opens its index.html, which loads its CSS, scripts and images by relative path. Links last from a minute to a year (a week by default) and keep working while the agent is stopped.

Turn every link of an agent off at once with Turn off all share links. Deleting the agent, or revoking the API key or agent token that made a link, ends it too.

Security and privacy#

• Each agent's sandbox is isolated from every other agent and customer.
• Shared pages are public to whoever has the link. Share only what you would publish.
• Each session gets its own API key, used for the agent's model and tool calls and revoked when the session stops. Its calls show up in your logs.
• Telemetry and session uploads to third parties are turned off.
• Signing out of the console signs you out of the agents you opened from it within a minute.

Limits#

• Up to 1,000 agents per account, 20 of them running at once.
• One session per agent at a time; up to 4 API runs at once per agent, one per session.
• A session lasts at most 24 hours; start it again to continue (the workspace is kept).
• No GPU.