Create Agent

Create Agent

Create custom OpenCode agents using markdown files with YAML frontmatter - the preferred method for readability and version control.

Markdown Format (Recommended)

Create ~/.config/opencode/agents/my-agent.md:

---
description: What this agent does
mode: subagent | primary
temperature: 0.3
permission:
  edit: deny
---

System prompt here...

File locations:

• Global: ~/.config/opencode/agents/*.md
• Project: .opencode/agents/*.md

Key Configuration

Common Patterns

Read-Only Research Agent

---
description: Research assistant with web access
mode: subagent
permission:
  read: allow
  webfetch: allow
  websearch: allow
  edit: deny
  bash: deny
---

Restricted Primary Agent

---
description: Planning mode without file changes
mode: primary
permission:
  read: allow
  edit: deny
  bash: deny
---

Code Review Subagent

---
description: Reviews code for quality and best practices
mode: subagent
permission:
  read: allow
  grep: allow
  glob: allow
  edit: deny
  bash:
    '*': ask
    'git *': allow
---

Prompt Design

Effective prompts include:

1. Clear role: "You are a..."
2. Responsibilities: "Your tasks are..."
3. Constraints: "You MUST NOT..."
4. Output format: "Structure responses as..."
5. Examples: 1-2 sample interactions
6. Tone guidelines: "Be concise and professional."

Tool Reference

Common permissions to configure:

• read / grep / glob - Code exploration
• edit / bash - File operations (deny for safety)
• webfetch / websearch - Internet access
• task - Invoking other subagents
• question - Interactive user prompts

Use allow, ask, or deny values. Use object syntax for granular control:

permission:
  bash:
    '*': ask
    'git status': allow
    'rm *': deny

Permissions

Control approvals in YAML:

permission:
  edit: ask # Prompt before edit
  bash:
    '*': ask # Ask all commands
    'git status': allow # Except git status

Permission Key for File Writes

Important: The edit key controls ALL file modification tools: edit, write, patch, and multiedit. There is no separate write: key — it is silently ignored if present. Always use edit: to restrict file writes.

Wildcard Path Patterns

Gotcha: ** is NOT a true globstar — it requires a literal / separator to work. Use *.md for files directly inside a directory; use **/*.md only when files are guaranteed to be in subdirectories.

permission:
  edit:
    "*": deny
    "~/.config/opencode/memory/*.md": allow    # correct — files directly in memory/
    # "~/.config/opencode/memory/**/*.md": allow  # WRONG — never matches flat files

Rule Evaluation Order

Last matching rule wins. Always put the "*" catch-all first, specific overrides after:

permission:
  edit:
    "*": deny            # catch-all first
    "~/safe/subdir/*.md": allow   # specific override last — this wins for matching paths
  bash:
    "*": ask             # catch-all first
    "git status": allow  # specific override last

external_directory and edit: Are Independent

Gotcha: external_directory only gates access to out-of-project paths. The edit: permission rules still run independently on top. You must configure both for agents that write to paths outside the project.

permission:
  external_directory:
    "~/.config/opencode": allow      # gates out-of-project access
  edit:
    "*": deny
    "~/.config/opencode/memory/*.md": allow  # still required — edit rules apply independently

Best Practices

• Start restrictive: Read-only by default
• Use edit: deny to block all file writes (not write: deny — that key doesn't exist)
• Use *.md not **/*.md for files directly inside a directory
• Configure both external_directory and edit: when writing outside the project root
• Specific descriptions: Help model know when to invoke
• Test incrementally: Verify tool access works
• Iterate: Refine prompts based on usage

Full Example

---
description: Security auditor that identifies vulnerabilities
mode: subagent
temperature: 0.2
permission:
  read: allow
  grep: allow
  glob: allow
  edit: deny
  bash: deny
---

You are a security auditor. Focus on:

- Input validation vulnerabilities
- Authentication flaws
- Data exposure risks
- Dependency vulnerabilities

Provide specific recommendations with code examples.
Never modify files - only report findings.

References

• Step-by-step guide
• OpenCode docs