OpenCode

OpenCode uses a combination of markdown files and a central opencode.json config file. Directives are stored as individual markdown files and also referenced in opencode.json, while MCP servers, permissions, ignore patterns, and settings all live in opencode.json.

Generated Files

dotai Source	Generated File
Directives	.opencode/instructions/*.md + opencode.json (instructions)
Skills	.opencode/skills/<name>/SKILL.md
Agents	.opencode/agents/*.md
Servers	opencode.json (mcp)
Hooks	Not generated
Permissions	opencode.json (permission)
Ignore patterns	opencode.json (watcher.ignore)
Settings	opencode.json (top-level keys)

Entity Details

Directives

All directives are always applied — OpenCode does not support conditional directive activation based on globs. Each directive is emitted as a markdown file under .opencode/instructions/ and its path is listed in the instructions array of opencode.json.

Agents

Agents are emitted as markdown files with YAML frontmatter under .opencode/agents/. In addition to the standard fields, OpenCode agents support the following additional properties: mode, temperature, top_p, steps, color, hidden, and disable.

MCP

MCP servers are written to the mcp section of opencode.json. Both HTTP and SSE transports map to OpenCode’s "remote" transport type. Tool filtering is not supported.

Permissions

OpenCode supports all three permission decisions: allow, deny, and ask. Permissions are written to the permission section of opencode.json using a nested format: tool name maps to an object of pattern-to-decision entries.

Ignore Patterns

Ignore patterns are written as an array to opencode.json under watcher.ignore.

Known Limitations

• No conditional directive support — all directives are always applied regardless of alwaysApply or globs settings.
• No hooks support.
• HTTP and SSE transports both map to "remote" — they cannot be distinguished in the output.
• No MCP tool filtering support.
• Project scope only.