pentect opencode

Pentect protects this OpenCode launch only. It does not edit opencode.json or change normal OpenCode launches. Background tasks and subagents use the same protected provider during the session.

Without --model, OpenCode's native picker exposes every provider Pentect can protect: OpenCode Zen, OpenAI, OpenRouter, Anthropic, and Google. /connect uses those native provider IDs, and each provider has its own local privacy gateway, so selecting or switching providers in the UI cannot bypass Pentect.

Provider setup and credentials stay under OpenCode's native provider ID. If opencode auth list shows a credential named pentect from an older Pentect release, reconnect the intended native provider once. Pentect does not guess a destination and copy an ambiguous stored credential automatically.

Run OpenCode's native authentication flow through Pentect before the first protected conversation when the provider requires an account:

pentect opencode auth login

Authentication commands retain OpenCode's complete provider list and do not install Pentect's temporary conversation routing. They exchange credentials, not conversation content. Providers not listed as supported below can be authenticated there, but protected conversations require a supported provider.

Normal local OpenCode arguments pass through. Attaching to another OpenCode server bypasses this process's gateway and is rejected. serve and web stay loopback-only; explicit non-loopback or mDNS exposure is rejected and inherited server configuration is narrowed to loopback for that launch. A model flag may appear anywhere before the explicit -- separator:

pentect opencode "Review this project" --model openai/gpt-5

After --, --model and -m are forwarded unchanged to OpenCode and do not select Pentect's provider route.

Pentect currently routes native opencode, openai, openrouter, anthropic, and google providers. Select one with OpenCode's provider/model form:

pentect opencode run --model openrouter/anthropic/claude-sonnet "Review this project"

An explicit --model enables only its matching provider for that protected launch. Without --model, the UI can switch among all supported providers; each remains bound to its matching local gateway.

Chat Completions is the default. Select Responses when the upstream supports it:

pentect opencode --api responses --model gpt-5

Session sharing and export

Protected launches disable OpenCode's external session sharing. The opencode export command is forwarded with OpenCode's own --sanitize option.

OpenCode's interactive /export command still creates a local transcript that can contain restored values. Pentect cannot sanitize that client-owned local file, so review it as sensitive data and do not publish it without inspecting it first.

Custom gateways

For Bifrost or another compatible gateway, keep its complete base path:

pentect opencode --model anthropic/claude-sonnet \
  --upstream http://127.0.0.1:8080/openai/v1

With a custom upstream, Pentect treats the model as an OpenAI-compatible gateway model ID. Custom endpoint variables and gateway credentials are advanced options; see Custom upstreams.

For custom upstreams, Pentect preserves model capabilities supplied in OPENCODE_CONFIG_CONTENT.provider["pentect-gateway"].models[MODEL_ID], including attachment, modalities, and token limits. If the custom model supports images, set PENTECT_OPENCODE_INPUTS=text,image for that launch; use text for a text-only model. Native provider routes keep OpenCode's model catalog capabilities. Without image capabilities, OpenCode may omit attachments before they reach Pentect, even when the upstream model itself supports vision.