09 Nov 2026 · ia, infrastructure · 5 min

Which companies let an agent operate your server today, and what they charge

Six companies now sell an agent that writes and runs code on your server or repository: Devin, Claude Code, Codex, Copilot, Cursor and Windsurf. Prices range from 8 to 200 dollars a month, and the billing model changes everything.

What “operate a server” means here

It is not autocomplete in the editor. It is a program that receives a task in natural language, opens files, runs commands, creates commits and, in some cases, opens a pull request on its own. The difference between them is where the agent runs (on your machine, in the company’s container or in the cloud), who holds the access key and how the account is billed.

Who charges what

Prices checked on November 9, 2026 on the official pages (Devin, Anthropic, GitHub, Cursor, OpenAI); they may change.

Teams: Copilot Business 19 USD per seat, Copilot Enterprise 39 USD, Copilot Max 100 USD; Devin Teams with an 80 USD monthly minimum and full seat at 40 USD; Cursor Teams from 30 USD per seat. Enterprise is quote-based everywhere.

What each one delivers differently

Devin and Windsurf run in the company’s cloud. You describe the task, the agent creates an isolated environment, executes it and returns a pull request. Access to your repository is through a GitHub token, not your SSH key. This is the most autonomous model and the one that consumes the most quota.

Claude Code runs in your terminal. It reads and edits local files, runs commands and calls MCP tools. The key stays with you; the risk is what the agent can do with your shell account. It is the most direct option for people who already work on the server.

Codex has two modes: in the browser and in the CLI/IDE, with cloud tasks and local sessions. The Plus plan covers moderate use; long tasks call for Pro.

Copilot is the cheapest to start and the most integrated with the editor. The cloud agent and agent mode in VS Code consume AI credits; code completion does not. A free tier exists, but it is for evaluation.

Cursor embedded the agent in the editor and charges through a weekly pool shared between Cursor and Grok. Third-party models (Claude, GPT, Gemini) are billed separately, at API price.

How to choose

Short task in the terminal, with you supervising: Claude Code or Codex CLI, on the 20 USD plan. A project that can run on its own for hours: Devin, at 20 to 200 USD depending on quota. Autocomplete and review in the editor: Copilot Pro at 10 USD or Cursor Plus at 8 USD. If you want to pay per token instead of per seat, Claude Code has an API-only route with no fixed monthly fee.

Aider, Cline and OpenHands are free and use your own API key. They are not companies charging a seat, but they solve the same problem for people who prefer to control cost.

What goes wrong

• Quota runs out mid-task: the agent stops or degrades to a smaller model. All plans have daily or weekly limits; only pay-per-use removes the cap.
• Too much access: an agent with your SSH key can delete or leak. Use a dedicated user, a separate key and minimum permission, as in harden SSH.
• Secret in the prompt: the agent may log the command. Pass the secret through a file or environment variable, never pasted into chat, as in env and secrets in a file.

The comparison as a table to copy

#!/usr/bin/env bash
cat <<'EOF'
Company              Entry (USD/mo)     Model
Cursor Plus          8                  weekly pool
Copilot Pro          10                 AI credits
Claude Code Pro      20                 shared quota
Devin/Windsurf Pro   20                 daily/weekly quota
Codex Plus           20                 included in ChatGPT Plus
Copilot Pro+         39                 7,000 credits
Codex Pro            100-500            three tiers
Copilot Max          100                20,000 credits
Devin/Windsurf Max   200                no daily cap
Claude Code Ultra    200                10x per week
EOF

FAQ

Which is cheapest to try? Copilot Free and Devin Free let you test without a card. Cursor and Windsurf have limited free plans. Claude Code has no reliable free tier; Pro at 20 USD is the entry point.

Can I run an agent with just my API key? Yes. Claude Code accepts per-token billing from 0.10 USD per million tokens on Haiku. Codex has an API route. Aider and Cline use your key directly and charge no seat.

Does the agent replace human review? No. All of them generate code that compiles and still gets things wrong. Review remains yours.

Which has the best cost for heavy daily use? For heavy use, compare the quota ceiling: Copilot Max at 100 USD gives 20,000 credits; Claude Ultra at 200 USD gives 10x Pro per week; Devin Max at 200 USD removes the daily cap. The choice depends on which model you use and how many tasks run in parallel.

Worth hiring for a production server? For occasional tasks, yes. For continuous operation, the cost and the risk of broad access weigh in; prefer an agent in an isolated environment with review before merge.

See also ready MCP servers, SSH agent and harden SSH.