Skip to content

REMOTE_LLMS.md

Latest commit

 

History

History
138 lines (83 loc) · 8.05 KB

REMOTE_LLMS.md

File metadata and controls

138 lines (83 loc) · 8.05 KB
Raw

Remote LLMs

WanGP can use Codex, Claude Code, or OpenCode as the shared LLM engine for Deepy and Prompt Enhancer. Remote engines run outside WanGP's local model runtime, so they do not occupy the VRAM reserved for media generation and do not require WanGP to unload a diffusion model before an LLM turn.

Remote engines require Deepy Prime because Prime exposes WanGP's MCP tools. The selected engine is also used for on-demand prompt enhancement and visual inspection. There are no separate engine selectors for those roles.

Privacy and credentials

Remote engines can send prompts, attached images, sampled video frames, and relevant tool results over the internet. Provider privacy, retention, and processing policies apply.

WanGP does not request or store provider passwords, API keys, access tokens, or refresh tokens. Authentication is owned by Codex, Claude Code, or OpenCode. The OpenCode custom-configuration field is saved verbatim, so do not put secrets in it; use OpenCode authentication or environment-variable references.

Common WanGP setup

  1. Open Configuration and select Prompt Enhancer / Deepy.
  2. Set Prompt Enhancer / Deepy LLM Engine to Codex, Claude Code, or OpenCode.
  3. Configure the selected engine as described below.
  4. Select Deepy Prime.
  5. Click Save Settings.
  6. Return to Media Generator, open Ask Deepy, and send a short test request.

Refresh queries the external engine and caches its model catalog in wgp_config.json. Refreshing does not change the selected model or reasoning effort. Automatic leaves the corresponding choice to the external engine or provider.

The chat footer displays token consumption when the engine reports it. The context limit is taken from the selected engine's model metadata; it is not a WanGP context-size setting. Remote providers manage their own context compaction. WanGP displays a compaction notice when the external protocol reports one, but it cannot display a provider-internal summary that was not returned.

Codex

Prerequisites and authentication

WanGP can use a standalone Codex CLI, an npm installation, or the compatible Codex binary bundled with the Codex VS Code extension. Leave Codex executable as codex for automatic detection, or enter the absolute executable path.

If authentication is required, send a Deepy request. WanGP will show the secure Codex sign-in link in the chat. Complete the browser sign-in, return to WanGP, and retry the request. The credential remains in Codex.

See the Codex CLI documentation.

WanGP fields

  • Codex executable: codex for automatic detection or an absolute path.
  • Codex model: the model reported by Codex; Automatic uses the recommended account default.
  • Reasoning effort: supported values depend on the selected model.
  • Refresh: queries the signed-in Codex account and caches the current catalog.

Claude Code

Prerequisites and authentication

WanGP needs both Claude Code and the compatible Python bridge. Install the bridge in WanGP's Python environment:

pip install claude-agent-sdk==0.1.40

Do not install an unpinned newer SDK into the WanGP environment because it can replace WanGP's MCP/Pydantic dependencies. WanGP detects Claude Code on PATH and compatible binaries bundled with VS Code, VS Code Insiders, Cursor, and Windsurf extensions.

Authenticate once with the configured executable:

claude auth login --claudeai

See the Claude Code authentication documentation.

WanGP fields

  • Claude executable: claude for automatic detection or an absolute path.
  • Claude model: an account model or alias; Automatic uses the account default.
  • Reasoning effort: sent only when explicitly selected and supported by the model.
  • Refresh: queries Claude Code and caches its model catalog.

OpenCode

OpenCode is the universal-provider option. WanGP talks to an OpenCode HTTP server and OpenCode talks to the selected provider. This makes providers such as OpenAI, DeepSeek, OpenRouter, local OpenAI-compatible servers, and others available without provider-specific WanGP backends.

Installation and provider authentication

Install OpenCode outside WanGP and authenticate providers using OpenCode. In the OpenCode terminal UI, run /connect, choose a provider, and follow its authentication flow. You can also use the OpenCode authentication commands documented by the project.

Provider credentials remain in OpenCode. After adding or removing a provider, restart an already-running OpenCode server if necessary and click Refresh in WanGP.

See the OpenCode provider documentation.

WanGP fields

  • OpenCode executable: opencode when it is on PATH, or an absolute executable path.
  • OpenCode server URL: defaults to http://127.0.0.1:4096.
  • OpenCode provider: providers currently configured and returned by OpenCode.
  • OpenCode model: models returned for the selected provider.
  • Reasoning effort: OpenCode model variant, when the model publishes variants.
  • OpenCode configuration: optional JSON/JSONC passed through OPENCODE_CONFIG_CONTENT only when WanGP starts a local server.
  • Refresh: queries /config/providers and caches the returned providers, models, context limits, defaults, and variants.

Refresh intentionally shows only providers currently available to OpenCode. It does not list every provider OpenCode supports. For example, to expose DeepSeek V4, connect the DeepSeek provider in OpenCode, supply a DeepSeek API key there, restart the server if needed, and refresh WanGP. DeepSeek V4 Pro and Flash then appear when reported by OpenCode's provider catalog.

Server lifecycle

WanGP first checks the configured server URL. If no server is reachable and the URL is local, WanGP starts opencode serve automatically on the configured port. The server remains alive between Deepy turns to preserve the OpenCode session and context.

  • Sending the first Deepy request starts a missing local OpenCode server.
  • Stop Deepy interrupts the current turn but keeps the server and session.
  • Reset clears the Deepy session and stops the OpenCode process only when that process was started by WanGP.
  • A manually started or otherwise pre-existing OpenCode server is never terminated by WanGP.
  • Before closing WanGP, use Reset if you want a WanGP-owned OpenCode process stopped cleanly.

To run the server manually:

opencode serve --hostname 127.0.0.1 --port 4096

Stop a manually started server with Ctrl+C in the same terminal.

Prompt Enhancer and visual inspection

Prompt Enhancer uses the same selected engine. Choose the normal on-demand prompt-enhancement action; no second remote-engine configuration is required.

When a remote engine is selected, Deepy visual inspection is routed to that remote engine. Images or sampled video frames may therefore leave the machine. The local inspection control is not exposed for this configuration.

Troubleshooting

Executable not found

An error mentioning WinError 2 means the executable name was not found on PATH. Enter its absolute path in the engine's executable field, save settings, and retry.

Refresh returns only a few OpenCode models

Run OpenCode's authentication listing command or /connect and confirm the desired provider is configured. OpenCode's catalog endpoint returns enabled providers, not its complete theoretical provider list.

An OpenCode model selection is ignored

Select an explicit OpenCode provider before choosing an explicit model. With both fields on Automatic, OpenCode uses its provider and model defaults.

Claude SDK dependency conflicts

Install exactly the pinned bridge version shown above in WanGP's environment. Do not add or upgrade WanGP's general requirements merely to obtain a newer Claude SDK.

Authentication or provider changes are not visible

Complete authentication in the external engine, restart an already-running external server when applicable, then click Refresh. Model catalogs are cached independently from the selected model.