Local vs cloud coding agents: where the work runs, and what that costs you

Updated 2026-09-02

Every autonomous coding agent has to put your repository somewhere it can run tests. Cloud agents put it in a VM the vendor operates and bill the work to the vendor's plan; a local runner keeps it on hardware you already own and bills the inference to the agent subscription you already pay for.

Denly is a control plane for coding agents (Claude Code, Codex, Cursor, Grok Build, Antigravity) that run on your own machines and your own subscription: you describe a task, the agent plans it, you approve, it builds locally and opens a pull request. That places it on the local side of the split described below, and this page is about what each side costs.

Our trade-off, first.Work runs only while your machine is awake and online: a task queued at 02:00 starts when the runner wakes. Two parties outside Denly still see your code, because the work needs them – your Git host receives what the runner pushes, and your AI provider receives the context its own CLI sends.

The one difference everything else follows from

A coding agent needs a checkout, a dependency install and a place to run tests. Whoever provides that machine sees the code and pays for the compute:

	Cloud agent	Runner on your own machine
Where the repository is cloned	A VM the vendor operates	The machine you already develop on
Who can read the source	The vendor, which holds the clone	Nobody but your machine. The control plane stores no file contents
Repository access granted to	The vendor, for the repositories you select	The control plane, for the repositories you select. It brokers a short-lived host token to the runner; the source itself never reaches it
Who pays for inference	The vendor's plan or credits	The agent subscription or API key already on that machine
Runs while your machine is off	Yes	No. Queued work waits for the runner to wake
Environment setup	Declared per project, rebuilt per task	Whatever your machine already has installed
Parallel capacity	The vendor's, so it scales past your hardware	Your hardware, plus one runner per machine you enroll
What breaks if the vendor has an outage	The agent stops	The dashboard stops; queued work resumes when it returns

The cloud agents, and the shape each one has

Agent	How it runs
OpenAI Codex cloud tasks	You connect GitHub or GitLab and choose the repositories Codex may access. Tasks run in isolated cloud environments in the background and come back with a diff and a pull request.
Devin (Cognition)	Given a ticket, it clones the repository on Cognition's infrastructure, edits, runs tests, iterates on failures and opens a pull request, with human review at gates.
Google Jules	An asynchronous agent built on Gemini. It clones the repository into a cloud VM, plans, and opens a pull request with the diff and tests.
Cursor Background Agents	Runs asynchronously either locally or on the web. In its cloud mode it starts an isolated Ubuntu VM, clones the repository and works on an `agent/` branch.

Product descriptions on this page were read from each vendor's own documentation on 2 September 2026. These products change; check their own pages before deciding.

Choose a cloud agent when

You want work to happen while nobody is at a machine, on a fixed nightly or weekend cadence.
You want more tasks in parallel than your hardware can hold.
Your project's environment is easy to declare and cheap to rebuild.
Handing repository access to the agent vendor is already acceptable at your company.

Choose a local runner when

You are already paying for an agent subscription and would rather not buy metered capacity twice.
Your source is not something you want cloned into a third party's VM. Granting a service repository access is not the same as handing it the files.
The project needs local state a fresh VM does not have: a database, a device, a licensed toolchain, a slow dependency tree.
You want the agent to run in the environment your tests already pass in.

What a control plane adds to the local option

Running the CLI yourself gets you the local half of the cloud agents' offer and none of the coordination: one terminal per project, and no record of who approved what. A control plane leaves the execution on your machine and moves the decisions to a dashboard, which is the arrangement this site is about. The runner polls outward, so no inbound port or tunnel is involved.

Questions people ask

Do cloud coding agents see my source code?

Yes. A cloud agent clones your repository into an environment the vendor operates, which is what lets it install dependencies and run your tests. Most teams accept that, but it is a grant of repository access to a third party, and it is the one thing a local runner does not ask for.

Can a cloud agent use my Claude or ChatGPT subscription?

It depends on whose agent it is. A first-party cloud agent meters against that vendor's own plan, so Codex cloud tasks draw on your ChatGPT plan. A third-party hosted service cannot reuse a CLI session stored on your machine, so it asks for an API key or resells inference at its own margin.

What happens to a local agent when I close my laptop?

Nothing is lost and nothing runs. Queued tasks wait, and the runner claims them when the machine wakes. If you need work to happen overnight, either leave the machine awake or use a cloud agent for that job.

What are the alternatives to Codex cloud?

Cloud alternatives with the same shape include Devin, Google Jules and Cursor Background Agents. The different shape is to run an agent CLI on your own machine and put a control plane in front of it, which is what Denly does for Claude Code, Codex, Cursor, Grok Build and Antigravity.

Is a local runner slower than a cloud agent?

It depends on what dominates your build. A local run skips the per-task VM boot and the dependency install, because the checkout and the caches are already on the machine; we have not published timings, so treat that as a description of the work skipped rather than a benchmark. A local runner is bounded by one machine either way, so a cloud agent wins when you want more parallel tasks than your hardware can hold.