Backends and Connection

Backends & Connection

The three backends

Backend	Typical role	How AgentStudio drives it
Claude	plan · review (or code)	the claude CLI (claude -p, structured output)
Codex	code · execute	the codex CLI (codex exec --json, resumable)
DeepSeek	plan, or write whole files	DeepSeek's OpenAI-compatible HTTP API

Any lane can run any backend — pick it from the backend dropdown in that lane's header.

Connection method: App login vs API key

Each backend has a per-lane toggle:

• App login — uses the CLI's existing OAuth/subscription login (claude / codex). Nothing to paste; it reuses how you already sign in.
• API key — paste a key and AgentStudio injects it into that backend only:
  • Claude → ANTHROPIC_API_KEY
  • Codex → OPENAI_API_KEY
  • DeepSeek → DeepSeek key (DeepSeek is always key-based; it has no app login)

App login and API key are independent per backend — e.g. Claude on App login while DeepSeek uses a key.

Connect / disconnect

• Connect — for App login this runs the CLI's login flow (browser) if you aren't already logged in; for API key it's "connected" as soon as a key is present.
• Disconnect — soft: it only forgets the connection inside AgentStudio. Your global claude / codex CLI login (shared with your terminal) is left untouched. For API-key mode, disconnect clears the stored key.

Connection state is session-scoped and not persisted — each launch starts disconnected by default, which is safe and predictable. Click connect once and AgentStudio detects an existing CLI login instantly.

Where keys live

API keys and the per-backend connection method are stored in the app's local settings file (settings.json in the app's userData directory), in plain text — same as most desktop dev tools. They are only sent to the corresponding provider.

Isolation from your machine

AgentStudio runs the CLIs with their own native memory turned off so they only use AgentStudio's memory — your personal CLI memory isn't pulled in or polluted. See Memory for details.