Cloud coding agents are easier to demo: hand over a ticket, get a pull request. Local agents are easier to trust: you watch the work, you see the diff, you keep the code on your machine. For most teams, the second property matters more than the first.
This page compares three local-first approaches we have reviewed — Aider, Zed, and OpenHands — and the cases where a hosted agent is genuinely the better choice.
Three ways to keep the agent local
The label covers different architectures:
- Aider is a terminal program that edits files and commits each change to git. The review surface is your git log. There is no editor integration to speak of, which is a feature if you already live in a terminal.
- Zed is a native editor with the agent inside it. The review surface is an inline diff in the buffer. It is the closest thing to pair programming, and the switch cost is real if your workflow depends on VS Code extensions.
- OpenHands is a self-hosted agent platform: a sandboxed runtime with shell and browser access, and a UI on top. The review surface is a session transcript with the commands it ran. It is infrastructure as much as it is an editor.
What local actually buys you
Three things, and they are worth naming precisely because they are usually described vaguely.
Access to your environment. A local agent can run the test suite that needs a database tunnel, call the service on localhost, and use the build cache. A cloud sandbox cannot, and the failure it produces looks like a logic error.
A data boundary you control. Repository content stays on the machine until it goes to the model provider. For teams with contractual constraints, that distinction decides the tool.
Diff-sized steps. Aider commits every change; Zed shows edits in the buffer. Both make it hard to produce a two-thousand-line change nobody reviewed. Cloud agents, by design, optimise for completing a task — and that means large diffs.
Where local agents are worse
Honesty matters more than advocacy here. Local agents are worse at three things:
| Weakness | Why it matters |
|---|---|
| Parallelism | One agent on one machine does one task, unless you run several sessions |
| Uptime | A long refactor competes with your own build for CPU and memory |
| Reproducibility | Two developers with different environments get different results |
A cloud sandbox solves all three by moving the work somewhere disposable. That is why the hybrid pattern — local for exploration, cloud for specified tickets — keeps showing up in teams that have tried both.
The setup that avoids the usual mess
Whichever you pick, spend an hour on the same four things before running a real task:
- A `.gitignore` that excludes secrets, plus a rule that the agent never reads `.env`. Aider and OpenHands both inherit whatever your shell can see.
- A test command that runs in under a minute. If the suite is slow, the agent stops verifying and starts guessing.
- A commit convention. Per-change commits are the review mechanism; if they are noisy, review stops.
- A working-directory scope. OpenHands makes this explicit with a sandbox; terminal agents need discipline.
When the cloud agent wins
If the task is already specified, the repository has a test suite that runs in the sandbox, and the work can be parallelised across tickets, a hosted agent is better. That is a real set of conditions, not a marketing claim.
The mistake is treating the choice as ideological. The question is not local versus cloud; it is who reviews the diff and when. Pick the architecture that puts the review where your team actually looks.
The switching cost, honestly
The reason local agents lose adoption battles is rarely capability. It is the switching cost of the environment around them.
Moving from VS Code to Zed means replacing extensions, keybindings, and the muscle memory of a decade. Adopting Aider means living in a terminal, which suits some workflows and disrupts others. OpenHands means running Docker and taking responsibility for a sandbox policy you did not have to write before.
None of those are technical failures. They are the cost of changing where you work, and they are paid by the whole team rather than the person who decided. That is why the honest way to evaluate a local agent is to try it for a week on real tasks, not to compare feature lists.
Frequently asked questions
Is a local agent more secure? It keeps repository content on your machine until it is sent to the model provider, which is a real boundary. It does not stop the model provider from seeing the code, and it does not protect against an agent reading secrets the shell can see.
Can I use Aider without an editor plugin? Yes, that is the design. It runs in any terminal, edits files directly, and commits to git. The editor is irrelevant to its workflow.
Does OpenHands require Docker? Effectively yes. The sandboxed runtime is the point of the product; running it without isolation removes the property that makes it trustworthy.
Last checked: 9 November 2026.
Related: Claude Code vs Codex vs Gemini CLI, how to evaluate a coding agent, and coding tools.