Can Droid run non-interactively?

Yes. Droid ships a headless style invocation where you pass the task as an argument and it exits with a status code. The flag set matters more than anything else here, because an unattended agent with no limits will happily loop all night. You want a task, a max-turn cap, and a timeout wrapping the whole command.

How do you wire Droid into CI?

The pattern below works on GitHub Actions, GitLab CI, and plain cron:

1. Install Node 18+ and run npm install -g @factory-ai/droid.
2. Export your model key as a secret, for example ANTHROPIC_API_KEY.
3. Call droid with the task and a turn limit: timeout 600 droid run "fix all ESLint errors in src/ and run tests" --max-turns 20
4. Fail the job if git diff --quiet returns false, so uncommitted changes are visible.

Keep the task narrow. “Fix all lint errors” is safe because the linter defines done. “Improve the codebase” is not, and the agent will meander into files you did not intend to change.

How do you control cost?

A CI run that touches 5-10 small files uses roughly 30k-80k tokens. At Claude Sonnet rates that is under $0.50. Two levers keep it there: a max-turns cap and a diff size check. If the produced diff touches more than, say, 15 files, reject the run instead of opening a pull request. This one rule prevents the classic failure where a bot rewrites half the repo overnight.

Where does it fit next to opencode and IDE agents?

If your team already scripts agents, opencode is the closer sibling — it was built to be embedded in shell pipelines. Droid’s edge is its planning loop on larger multi-file tasks, which shows up when a single lint rule needs changes across many modules. For interactive day-to-day work, stick with the normal Droid session; reserve headless mode for repetitive jobs.

What are the limits?

Unattended agents cannot ask clarifying questions. If the task is ambiguous, you get a plausible but wrong diff, and someone reviews it anyway. There is also no built-in dedup: run it twice on the same branch and you may get two competing patches. Gate every run behind a fresh branch and a human merge.

Quick Summary

• Droid runs headless with a task argument and an exit code.
• Always set a max-turn cap and an external timeout wrapper.
• Typical CI run: 30k-80k tokens, under $0.50.
• Reject any diff touching more than ~15 files.
• Headless mode cannot ask questions — use it only for well-defined jobs.

Last updated September 2026.

What is Droid CLI?

Droid is an agentic developer tool that lives in your shell rather than an IDE. It scans the current directory, builds a plan, then writes code and executes shell commands with your approval. Factory ships it as a Node package, so Node 18 or newer is required. Unlike a chat window, Droid keeps the working tree in context, which is why it handles multi-file refactors without you pasting file contents.

How do you install Droid?

Run one command:

• npm install -g @factory-ai/droid
• Then start it inside a project: cd my-app && droid

On first launch it asks for credentials. You can sign in with a Factory account or supply your own API key for Anthropic, OpenAI, or a local endpoint. The config lands in ~/.factory/config.json, so switching models later is a two-line edit rather than a reinstall.

How do you run your first task?

Open a repo you know well, then type a request that names the file and the outcome. For example: “Add input validation to the signup handler in api/auth.js and return 400 on a missing email.” Droid proposes a diff before touching disk. Press y to apply, n to reject, or type a correction to revise the plan.

Three habits make it reliable:

1. Start on a clean git branch. Droid edits real files, and git diff is your undo button.
2. Ask for one change at a time. Bundled requests produce bigger diffs and more review work.
3. Let it run tests. Say “run the test suite and fix failures” and it will loop until the suite passes or it gets stuck.

Is Droid free?

The Droid CLI itself is free to install. You pay for the model behind it. A mid-sized refactor typically burns 20k-60k tokens, which on Claude Sonnet pricing lands around $0.10-$0.40 per task. If you already have an API key, there is no separate subscription. Teams that want centralized billing and shared context can use Factory’s paid plan, but solo developers can run it indefinitely on their own keys.

Droid vs other terminal agents

Droid competes with opencode and IDE-based tools like Cline. Droid leans toward autonomous multi-step work, while opencode stays lighter and more scriptable. If you mostly want inline autocomplete, neither beats an IDE plugin. Droid is best for people who already live in tmux and want an agent they can pipe into scripts.

Common pitfalls

Two things trip up new users. First, Droid can overwrite a file if you approve a diff without reading it, so glance at the hunk headers. Second, long tasks drift: if the agent starts editing unrelated files, stop it and restate the goal with a narrower scope.

Quick Summary

• Droid is Factory AI’s terminal coding agent, installed via npm install -g @factory-ai/droid.
• It needs Node 18+, edits files directly, and asks for approval before each change.
• The CLI is free; you pay only for model tokens, roughly $0.10-$0.40 per task.
• Works best on a clean git branch with one request at a time.
• Good fit for terminal-first developers; overkill if you just want autocomplete.

Last updated September 2026.