Skip to main content
Version: 2.0.0

Providers & Login

ClaudusBridge ships one embedded coding agent that can drive three different AI providers using your own accounts:

  • Claude (Anthropic)
  • ChatGPT / Codex (OpenAI)
  • Gemini (Google)

There are no API keys to paste and no per-token billing from the plugin. You sign in with the same account you'd use in each provider's own app, and that provider handles login and model requests. The native CLI and local MCP bridge can also be used independently of the embedded agent.

One agent, three providers​

The embedded agent is ClaudeCodeAntigravity (CCAG) — a Claude Code build that also speaks to ChatGPT (Codex) and Gemini in the same session. It ships inside the plugin as a single, self-contained Node bundle at Resources/Agent/:

  • cli.js — the bundled agent (Node target, vendor-free, ~30 MB).
  • cli-node.js — a tiny ESM entry that imports cli.js.

Because the bundle is vendor-free, the plugin needs no external agent install. The only host requirement is a system Node.js install, which the plugin uses to run the bundle.

The agent is reachable two ways, both local:

  • In-editor terminal. The ClaudusCode module hosts a Slate terminal (SClaudusCode) backed by a real PTY (ConPTY on Windows, forkpty on Mac/Linux) and a VT parser. This is the interactive, type-to-the-agent experience.
  • Programmatic ACP. FClaudusAcpClient spawns the bundle headless via node "cli-node.js" --acp and drives it over the Agent Client Protocol (JSON-RPC over stdio) — no PTY required. Console commands: ClaudusCode.Acp.Start, ClaudusCode.Acp.Stop, ClaudusCode.Acp.Prompt.

How sign-in flows​

Sign-in happens through the agent, using each provider's normal OAuth login — the same browser-based "Sign in with…" flow you already know:

  1. Start the agent (open the in-editor terminal, or start an ACP session).
  2. Trigger the provider's login from the agent (its standard login command/flow).
  3. The provider's OAuth page opens in your browser. Approve it with your own Claude, ChatGPT/Codex, or Gemini account.
  4. The session token is stored by the agent locally on your machine. From then on the agent uses your account's normal entitlements.

Because this is OAuth against the provider, the plugin never sees or stores an API key. There is no ClaudusBridge billing layer in the path — your usage is just your account's usage.

Switching models and providers​

Once signed in, you switch between providers and models the same way you would in the underlying Claude Code-style agent — through its model/provider selection inside the terminal or via the agent's own commands. The single agent process can target Claude, ChatGPT/Codex, or Gemini; pick whichever account you're signed into and whichever model that account exposes.

Over ACP, the same agent applies a permission mode to each session, which governs how much the agent may do on its own:

ModeBehavior
defaultAsks for permission on sensitive actions.
acceptEditsAuto-accepts edits, still prompts for riskier actions.
bypassPermissionsRuns without permission prompts.
planPlans without making changes.

For ACP, these modes are supplied to the agent. If it sends a permission request back to FClaudusAcpClient, the host delegates the decision to its bound handler; with no handler, it cancels the request. There is no terminal prompt on this programmatic path. See ACP Drive Path for binding your own permission UI.

Local bridge and provider connections​

The agent bundle runs locally under Node. External MCP clients talk over stdio to the Claudus CLI, which authenticates to Unreal over loopback. ClaudusBridge 2.0 does not expose an HTTP MCP server.

Provider login and model requests use the selected provider's network service. Other agent tools, external integrations and Concert sessions may have their own network behavior. A local bridge is not a promise that every action an AI client performs stays on the machine.

Scripted claudus commands do not require Node or an AI account. The requirements below apply to the embedded agent.

Requirements​

  • Unreal Engine 5.8 (Windows / Mac / Linux).
  • A system Node.js install on PATH (used to run the embedded agent bundle).
  • An account with at least one of the providers — Claude (Anthropic), ChatGPT/Codex (OpenAI), or Gemini (Google).

Tip: verify the bundle runs against your Node install with node cli-node.js --version from Resources/Agent/ — it should print the agent version (e.g. Claude Code).