Skip to main content

OpenCode harness

OpenCode is the other main native-harness Guard path: Guard handles artifact trust and provenance while OpenCode keeps its own permission model for tool execution.

What Guard detectsDirect link to What Guard detects​

The OpenCode adapter reads:

  • ~/.config/opencode/opencode.json
  • <workspace>/opencode.json
  • <workspace>/opencode.jsonc
  • <workspace>/.opencode/commands/*.md

Inside config, Guard inspects the mcp block and records one artifact per declared server. For workspace commands, Guard records one artifact per Markdown command file in .opencode/commands.

Approval modelDirect link to Approval model​

OpenCode uses the native-harness tier:

  • Guard authors artifact-level policy before launch
  • OpenCode keeps its native allow or deny behavior for tool execution
  • the goal is package-level trust without replacing OpenCode's built-in permission semantics

Launch behaviorDirect link to Launch behavior​

OpenCode shares one special launch rule with Claude Code: when you run it from a workspace, Guard passes the workspace path through on launch so OpenCode opens in the right directory.

Use this loop:

hol-guard detect opencode --json
hol-guard install opencode
hol-guard run opencode --dry-run
hol-guard run opencode

What makes OpenCode differentDirect link to What makes OpenCode different​

OpenCode is the only current Guard harness that combines:

  • JSON or JSONC config discovery
  • MCP server detection from a command list
  • workspace command discovery from .opencode/commands/*.md

That makes it a good fit when you want Guard to watch both runtime server wiring and workspace-authored command surfaces.

hol-guard detect opencode --json
opencode --help
hol-guard run opencode --dry-run
hol-guard receipts
hol-guard status

If your workspace uses .opencode/commands, make one small change to a command file and confirm Guard records it as a changed artifact before the next launch.

Next guidesDirect link to Next guides​