Skip to content
The Test Cabinet The Test Cabinet

Overview

OpenCode (slug opencode) is an open-source coding agent CLI, driven non interactively through its opencode run subcommand. It reaches its model through OpenRouter, so an OpenCode run reports OpenRouter-style model IDs and is priced from OpenRouter’s listed rates. See the OpenCode site for the harness itself.

Model IDs

Section titled “Model IDs”

OpenCode runs models through OpenRouter, so its model IDs are OpenRouter slugs carrying an openrouter/ prefix. Examples (illustrative, not exhaustive):

  • openrouter/minimax/minimax-m3
  • openrouter/google/gemini-3.5-flash

Invocation

Section titled “Invocation”

The harness probes and invokes the opencode binary. It is installed into the run container at run time with:

Terminal window
npm install -g opencode-ai && npm cache clean --force

A session is run with opencode run in non-interactive mode. The Test Cabinet passes these flags:

FlagPurpose
--format jsonEmit the line-delimited JSON event stream consumed for events and usage.
--dangerously-skip-permissionsRun unattended, without per-action approval prompts.
--model <id>The model to run.

The prompt is passed as the final positional argument.

Authentication. OpenCode’s API key is sourced from OPENROUTER_API_KEY on the host and passed straight through to the run container under the same name.

Pricing. OpenCode’s model IDs already carry an openrouter/ prefix, so the comparable-cost lookup strips it before consulting OpenRouter — openrouter/minimax/minimax-m3 becomes minimax/minimax-m3. OpenCode does not self-report a run cost, so the comparable cost is always OpenRouter-derived; see Metrics.

See Events for how OpenCode’s output is normalized and Metrics for how its usage is counted. For the harness layer these pages fit into, see Harnesses.