Data and privacy

Conversations and settings are saved locally. To answer a request, fx sends the relevant context to your selected provider. Networked tools send data to the services they use.

Model request data

Each model request includes your prompt and the context assembled for that turn. Context can include conversation history, applicable AGENTS.md and skill instructions, attached images, and file or tool content already loaded into the session.

fx does not automatically package or upload your workspace, Git history, session files, traces, or recordings. That data can still leave the machine when it is loaded into model context or sent by a networked tool.

Context compaction sends older prompts, responses, tool calls, and tool results to the current session model for summarization. On AI Gateway, when that model can't write the summary, fx can send the same request to a model from another provider family: anthropic/claude-sonnet-5, or openai/gpt-6-sol when the session uses an Anthropic model. Reviews in auto mode, vision fallback, and generated session titles can also make additional model requests. Titles use up to 2 KiB of the first prompt; disable them with session_titles: false. While a Claude conversation through AI Gateway waits, cache warming resends its last request for up to 30 minutes; disable it with cache_warming: false. See Additional model requests.

Where requests go

Your selected connection determines where model requests go:

ConnectionWhere fx sends model context
CodexThe service associated with your ChatGPT subscription
GrokThe service associated with your Grok subscription
AI GatewayGateway, which forwards the request to the selected model provider
Custom connectionThe endpoint configured in base_url

Remote requests are subject to the selected service’s data controls and retention policy. MCP servers and web search send requests to their own services.

Reviews in auto mode send the pending action and context from the conversation to the selected connection's reviewer. With an AI Gateway connection whose review_model is typesafeai/jev, fx sends reviews to TypeSafe directly when TYPESAFE_API_KEY is set, and through Gateway otherwise.

Local credentials and sessions

fx stores settings, saved sign-ins, conversations, prompt history, a cache of the usage your provider reports, MCP configuration, and installed skills under ~/.fx/, or under the folder FX_HOME names when it is set. fx creates that folder with access for your user account only. Opt-in debug logs and recordings also go there by default. To stop saving prompt history, set prompt_history.enabled to false.

On macOS, your Vercel sign-in and a saved AI Gateway API key live in Keychain, which FX_HOME does not change. On Linux, they are the files ~/.fx/auth.json and ~/.fx/api-key, with 0600 permissions. Setting FX_DISABLE_KEYCHAIN=1 keeps the Vercel sign-in in ~/.fx/auth.json on macOS too, and turns off stored API keys on every platform. ChatGPT and Grok sign-ins are always files: ~/.fx/chatgpt-auth.json and ~/.fx/grok-auth.json. MCP OAuth credentials also use Keychain on macOS when available; see OAuth configuration for the file fallback.

Session files remain local, but fx sends the relevant conversation context again when you continue a session. Use fx ask --no-save when a one-off request should not create a session.

Product telemetry

fx does not send product telemetry or usage analytics to a separate fx service. /usage and fx usage read usage from your selected provider with your sign-in, and /trace assembles diagnostics locally.

Update checks

Automatic updates are on by default. Native fx reads static release metadata from releases.fx.sh shortly after startup and every 30 minutes until an update is ready. The request contains no fx-generated machine or installation identifier. To turn automatic updates and their checks off, set auto_upgrade to false in ~/.fx/settings.json, or set FX_AUTO_UPGRADE=0.

Local inference

Use a custom connection to send model requests to a local server such as Ollama. Update checks, web search, and remote MCP servers can still contact external services.

Sharing diagnostics

/trace writes a diagnostic report to a local temporary file and, on macOS, copies the file to your clipboard. /feedback opens the feedback form without uploading diagnostic data.

To share a trace as a link, fx.sh/share redacts and encrypts it in your browser and uploads only the encrypted text. The decryption key stays in the part of the link after #, which your browser never sends to fx.sh. Anyone with the link can read the redacted trace until the link expires, 7 days after you create it.

Review and redact prompts, code, paths, commands, model output, and secrets before sharing a trace or recording. See Share feedback for details.

AI Gateway controls

For a Gateway connection, account and team settings apply to requests from fx. See Gateway's Zero Data Retention settings for its routing and retention controls. Those settings apply to Gateway requests, not to separate Codex or Grok connections.