FAQ

The questions that come up most often.

Is agentry actually free?

Yes. agentry the service costs nothing. No trial, no token meter, no usage cap.

You pay your model provider (Anthropic, OpenAI, OpenRouter, whoever) directly for tokens, and your server provider (or your electric bill, if you're using your laptop) for the compute. agentry takes no cut.

Why do you need my server? Why not host it yourselves?

The short answer: your data, your code. We're a routing layer, not a hosting provider. Apps you build with agentry are deployments running on the machine you connected — same model as if you'd Dockerized the app yourself, with agentry handling the public URL and the connection.

If you'd rather we hosted it, well — that's the dominant pattern in the rest of the market. agentry exists because that pattern has tradeoffs (vendor lock-in, your data lives on the vendor's servers, you pay per token for the AI plus per request for the hosting).

Can I use it without a credit card?

Yes. agentry never asks for a card. The model provider you choose might (Anthropic, OpenAI, etc. require payment for serious usage). Some — Google's Gemini free tier, OpenRouter's free models — let you start without one.

Which AI provider should I pick?

For prototyping: OpenRouter if you want to A/B models without rewiring. Single key, every provider, drop-in OpenAI API shape. The small markup pays for itself the first time you switch model strings without redoing setup.

For production: directly with your favorite provider to avoid the markup. Anthropic for code-heavy work, OpenAI for the broadest model selection, open-weights for cost control or privacy.

For free experimentation: OpenRouter's free models (DeepSeek, Llama variants) or Google's Gemini Flash free tier.

Can I use Claude in my harness and GPT in my app?

Yes, that's the common pattern. The harness's model is configured in the harness; the model your deployed app calls server-side is configured by binding an AI provider in agentry. They're independent — drive Claude Code with Claude, run GPT-powered features in the app it builds.

What languages / frameworks does agentry support?

Anything agentry can build automatically: Node.js (Next.js, Vite, Astro, SvelteKit, plain Express), Python (FastAPI, Django, Flask), Go, Ruby (Rails), Rust, PHP (Laravel), Elixir (Phoenix), Java, .NET, Bun, Deno.

The sandbox runtime image is Linux + Docker, so you can install anything that runs in a container. If agentry doesn't auto-detect your stack, drop a Dockerfile in the project root — agentry uses it.

Can two sandboxes share a database?

They can — and by default they do. A binding hands every sandbox the same connection URL, verbatim, so two sandboxes pointed at the same bound database see the same data.

If you'd rather keep prototypes apart, that's a code convention, not an automatic split: each sandbox gets a unique AGENTRY_APP_NAME, and a well-prompted agent uses it as the database name / Postgres schema / Redis key prefix (client.db(process.env.AGENTRY_APP_NAME)). Same code in production uses the deployment slug, so dev and prod stay separate too.

To force sharing regardless, have the code use a fixed name instead of AGENTRY_APP_NAME — e.g. hardcode client.db("shared"), or pin a database in the URL:

agentry env set --sandbox <sandbox-id> MONGODB_URL "mongodb://...?database=shared"

Can my app outlive agentry?

Yes. Every deployment is a standard Docker container running on your server. If agentry the service went away tomorrow:

• The container keeps running.
• The the bridge URL stops working (because that's agentry's edge).
• You can keep serving the container yourself — docker run it directly, put nginx in front, point your own DNS at it.

There's no agentry-specific code baked into your container image. The Dockerfile (or the one agentry generates for you) is plain Docker.

Can the agent install OS packages?

In a sandbox, yes. The sandbox runs as a non-root user with sudo for apt. Common pattern:

agent: command_run "sudo apt-get install -y libpq-dev"

For things that need different OS layers entirely, write a Dockerfile and agentry uses it at deploy time.

What's the difference between a sandbox and a deployment?

	Sandbox	Deployment
Purpose	iterate, prototype, develop	run for real users
State	mutable filesystem, shell access	immutable container image
Lifetime	as long as you keep it	until you stop / delete
URL	preview URL on *.agentry.live	stable production URL
Auth	open in dev	configurable (public / org-only)

You build in sandboxes. You ship deployments.

Why does the agent sometimes create a sandbox with a suffix?

If you (or the agent) request a sandbox name that already exists, agentry auto-allocates a suffixed name (my-app-7f2a) to avoid overwriting the existing one. The returned sandbox_id will be the suffixed version — use it for all subsequent tool calls.

This is the same protection that keeps two parallel chats from clobbering each other's work.

Can I run agentry without a harness?

Technically yes — agentry is a regular CLI. You can agentry sandbox create, agentry command run, etc., from a terminal. But the value of agentry is letting an AI do the work for you; without a harness you're driving the sandbox manually.

If you want to script agentry operations (CI, automation, batch jobs), the CLI + REST API both work. See the CLI reference.

How do I move data between sandboxes?

Tar it out, tar it in:

agent_1: command_run "tar czf /tmp/data.tar.gz /workspace/projects/app/data"
agent_1: file_read /tmp/data.tar.gz   # bytes come back to your harness
# share with agent_2 somehow (paste, save, transfer)
agent_2: file_write /tmp/data.tar.gz <bytes>
agent_2: command_run "tar xzf /tmp/data.tar.gz -C /"

In practice: more pleasant to use git (push from one sandbox, pull in the other) or to scp directly between sandboxes if they're on the same server.

How do I get help?

Two paths:

• Search these docs. Use the search bar at the top of every page.
• Email support@agentry.run for stuck cases. Include the deploy URL, sandbox id, or error message — much faster than describing the problem from memory.

For a faster broadcast — bug reports, feature requests, security disclosures — see the appropriate page in Reference or email the security team at security@agentry.run.

Why no Discord / Slack / community forum?

We're focused on shipping the product right now, not running a community. We'll know when the time is right to start one. In the meantime, the email path is faster anyway — engineers read it.

Next

• Get started — start at the beginning.
• Quick start — five-minute path to a live app.
• Concepts — the mental model.