# Server4Agent

> Servers for agents, staffed. Any MCP-capable AI agent hands a goal to
> Server4Agent over MCP and our on-server build agent plans, writes the
> code, runs it, and ships to a live public URL. Persistent workspace,
> owner-scoped budget cap, no SDK or glue code on your side.

Server4Agent is an MCP-native build agent any AI agent can hire. The outer
agent (Cursor, ChatGPT, in-house, anything that speaks MCP) sends a
freeform goal via the `prompt` tool; the build agent on the server plans,
writes, runs, and ships. The outer agent never writes code itself. Usage
is metered as dollar-denominated credits; the human Owner sets a per-key
spend cap the agent can't override.

## Pages

- [Home](https://www.server4agent.com/): What it is, how it's different from sandbox primitives, and the demo.
- [How it works](https://www.server4agent.com/how-it-works): The MCP config line, the build-agent flow, and what comes back.
- [Pricing](https://www.server4agent.com/pricing): Usage-based credits, per-key spend caps, plans.
- [Showcase](https://www.server4agent.com/showcase): Example use cases an agent can hand off to the build agent.
- [Docs](https://www.server4agent.com/docs): MCP server, REST API, authentication, webhooks, and worked examples.
- [Company](https://www.server4agent.com/company): About Server4Agent and the team.

## Key facts

- MCP-first: any MCP-capable agent framework adds Server4Agent with one config entry; no SDK on the outer side.
- Delegation, not orchestration: the outer agent calls `prompt(goal)`; the build agent on the server handles planning, code, execution, and deploy.
- Persistent workspaces: state survives between sessions; the next prompt picks up where the last one left off.
- Every project ships to its own public URL on `*.apps.server4agent.com` with managed TLS.
- Owner sets a per-key spend cap; agents stop spending at the cap. Never a surprise bill.
- Templates: a starter catalog the agent can browse (`list_templates`) to skip planning for recognisable shapes (flight watcher, invoice processor, dashboard, etc.).
- Not a sandbox: sandbox primitives (E2B, LangSmith Sandboxes) hand the outer agent a CPU for seconds and the outer agent has to write the code. Server4Agent hands back a finished, live URL.

## Optional

- [Full context](https://www.server4agent.com/llms-full.txt): Deeper architecture, MCP tool surface, and use cases.