Tably Team
Home › Blog › Costs and security
Costs and security

Security and privacy in Tably Team: what happens with your API keys, your code and your data

Where your keys are stored, what each AI provider sees, where your code lives, what data leaves your machine, how the subscription token works, and what options you have if nothing can leave your infrastructure.

4 min readby the team at Tably TeamsecurityprivacyAPI keys

Before pasting an API key into any tool, it is worth knowing what it does with it. This guide explains how Tably Team handles keys, code and data, without euphemisms.

Your API keys

  • They are stored encrypted in your Tably account, on the server where your instance runs. They are never written into your projects’ code or into the logs.
  • They are used only for your requests: when a role on your team or your builder calls the provider, it does so with your key. No other user of the instance sees or uses them.
  • You can view the last characters, replace them and delete them in “Models and API keys”. Deletion is immediate.
  • If you run Tably on your own machine, the keys never leave it except to go to the corresponding provider.

Recommendation: create keys specifically for Tably with each provider, with a spending limit, so you can revoke them without affecting anything else.

The Claude subscription token

To build with your Claude subscription on a server (for example, Tably deployed in the cloud), you generate a token with claude setup-token and paste it into your settings.

  • It is stored encrypted in your account, like an API key.
  • It is used only for your builds, with a Claude Code environment isolated per user.
  • You can revoke it from Tably (by deleting it) and from your Anthropic account.

It is equivalent to signing in to Claude Code on another computer of yours.

What each AI provider sees

When a role uses a model, the provider receives what that role needs: the brief, the debate context, the relevant project files, the reference descriptions. This is unavoidable: the model has to read in order to work. What each provider does with that depends on its policies:

  • The paid APIs of Anthropic, OpenAI, Google and most providers do not train on your data by default. Check each one’s current terms.
  • Some providers (DeepSeek, Moonshot) process data in different jurisdictions. If that matters, use the same models via OpenRouter choosing providers that do not retain data, or run local models.
  • With Ollama, nothing leaves your machine.

Tably lets you choose the provider role by role precisely so you can make that decision per project.

Your code

  • Each project lives in your account, in its own git repository.
  • You can download it, clone it and push it to your GitHub at any time.
  • No other user of the instance has access to your projects.
  • If you delete a project, its repository, its files and its history are deleted.

Reference apps

When you provide reference URLs, Tably visits them with its own browser. The screenshots and descriptions stay inside your project. With reverse engineering enabled, the bundles and public responses that are analyzed stay there too. Nothing is shared with other users.

The apps you publish

  • Each published app has its own environment variables and secrets, encrypted, which no other app sees.
  • The Neon, Supabase, Cloudflare or Fly credentials you enter for Tably Cloud are stored encrypted in your account and used only to provision your projects.
  • Your app’s service keys (payments, email) are entered as project secrets, never in the code.

If nothing can leave

Tably can run entirely on your infrastructure: your machine or your server, with your database. With Ollama for the models and Claude Code or Codex via subscription for building, the full workflow works without entering a single third-party API key. There is a deployment guide (Docker and Railway) in the project repository.

What Tably does not do

  • It does not read your projects to train anything.
  • It does not share your keys, tokens or code with third parties, except the AI provider you chose for each call.
  • It does not charge hidden fees for AI usage: the cost you see is the providers’ cost.

Best practices

  1. Keys per service with a spending limit.
  2. One provider per project according to the sensitivity of the data.
  3. Review the project secrets before publishing.
  4. Delete the subscription token from servers you no longer use.

In summary

Your keys are stored encrypted and used only for you; your code is yours and portable; what each AI provider sees is your decision, role by role; and if necessary, everything can run on your infrastructure.

Keep reading

Costs and security

How much it costs to build an app with AI: tokens, subscriptions and what actually drives the bill

An honest breakdown of the cost of building an app with an AI team: what providers charge, what a council debate costs, how to build without spending tokens using a subscription, and how Tably shows the real cost.

4 min readRead →

Ollama: run AI models on your own computer, free and in private

What Ollama is, which models you can run locally (Llama, Mistral, Qwen, DeepSeek, Gemma), what hardware you need, what quality to expect and how to use it on an AI team without any data leaving your machine.

4 min readRead →

Mistral: the European model, what it is good for and what role to give it on your team

A guide to the models from Mistral AI. Its European focus on privacy, its Small, Medium and Large variants, its performance on code and the roles on an AI team where it contributes most.

2 min readRead →

Try Tably Team with your idea

Describe what you want to build and watch the team debate it, design it and build it.

Create a free account