OpenCode

OpenCode is the open-source omnichannel agentic coding CLI. As a Subshell harness it runs the interactive TUI in a pane, with the subshell MCP server wired into every launch through a merged config layer, so your own OpenCode configuration keeps working untouched.

Finding the binary

Subshell resolves opencode on the machine that will run the pane: the opencode on the PATH there (including what a login shell would find), then the installer's own location under ~/.opencode/bin. Set OPENCODE_PATH on that machine's environment to name the binary explicitly. If detection finds nothing, Installing Agent CLIs covers the one-line installer, and the vendor CLI docs the manual path.

What a launch looks like

A bare OpenCode launch opens the TUI, so that is what a subshell runs. Settings arrive as per-invocation flags: the preset editor offers model (-m, in provider/model form), agent, and auto-approve permissions, plus the flags and environment variables OpenCode users tend to want. OpenCode has no create-time flag for naming its session, so the subshell's display name is not forwarded to it; Subshell's own naming, sharing, and logs still apply.

Cross-subshell tools, automatically

The subshell MCP server (channels, sibling inspection, coordination) is registered for every OpenCode subshell automatically, using OpenCode's own config-merge behaviour: OpenCode loads any config file named by OPENCODE_CONFIG and merges it between its global and project configs.

So Subshell composes a tiny per-subshell config layer holding just the subshell MCP entry and points OPENCODE_CONFIG at it. The control plane builds it at launch; on a node, the launch command carries the layer to that machine, where it lands beside the pane and names that machine's own subshell binary as the server. Either way the result is the same: the entry is added, your global and project config are preserved, and every MCP server you configured yourself is still there.

Because the registration names only the command to spawn, the subshell mcp child inherits the credential baked into its pane's environment, each subshell's own scoped API key, so no secret ever sits in the config layer.

See also

• Codex: the same automatic wiring, done with per-invocation overrides instead of a config layer
• Presets: saved per-harness configuration, including OpenCode's settings fields
• The MCP server: what the registered tools actually give the agent
• Installing Agent CLIs: the in-page installer for the opencode binary

Edit on GitHub