opencode/packages/web/src/content/docs/agents.mdx at 87a196687a8ae5adc3e5655a7b402b139d9fe2b7 · anomalyco/opencode

title	Agents
description	Configure and use specialized agents.

Agents are specialized AI assistants that can be configured for specific tasks and workflows. They allow you to create focused tools with custom prompts, models, and tool access.

:::tip Use the plan agent to analyze code and review suggestions without making any code changes. :::

You can switch between agents during a session or invoke them with the @ mention.

Types

There are two types of agents in OpenCode; primary agents and subagents.

Primary agents

Primary agents are the main assistants you interact with directly. You can cycle through them using the Tab key, or your configured switch_agent keybind. These agents handle your main conversation. Tool access is configured via permissions — for example, Build has all tools enabled while Plan is restricted.

:::tip You can use the Tab key to switch between primary agents during a session. :::

OpenCode comes with two built-in primary agents, Build and Plan. We'll look at these below.

Subagents

Subagents are specialized assistants that primary agents can invoke for specific tasks. You can also manually invoke them by @ mentioning them in your messages.

OpenCode comes with two built-in subagents, General and Explore. We'll look at this below.

Built-in

OpenCode comes with two built-in primary agents and two built-in subagents.

Use build

Mode: primary

Build is the default primary agent with all tools enabled. This is the standard agent for development work where you need full access to file operations and system commands.

Use plan

Mode: primary

A restricted agent designed for planning and analysis. We use a permission system to give you more control and prevent unintended changes. By default, all of the following are set to ask:

file edits: All writes, patches, and edits
bash: All bash commands

This agent is useful when you want the LLM to analyze code, suggest changes, or create plans without making any actual modifications to your codebase.

Use general

Mode: subagent

A general-purpose agent for researching complex questions and executing multi-step tasks. Has full tool access (except todo), so it can make file changes when needed. Use this to run multiple units of work in parallel.

Use explore

Mode: subagent

A fast, read-only agent for exploring codebases. Cannot modify files. Use this when you need to quickly find files by patterns, search code for keywords, or answer questions about the codebase.

Use compaction

Mode: primary

Hidden system agent that compacts long context into a smaller summary. It runs automatically when needed and is not selectable in the UI.

Use title

Mode: primary

Hidden system agent that generates short session titles. It runs automatically and is not selectable in the UI.

Use summary

Mode: primary

Hidden system agent that creates session summaries. It runs automatically and is not selectable in the UI.

Usage

For primary agents, use the Tab key to cycle through them during a session. You can also use your configured switch_agent keybind.
Subagents can be invoked:
- Automatically by primary agents for specialized tasks based on their descriptions.
- Manually by @ mentioning a subagent in your message. For example.
```
@general help me search for this function
```
Navigation between sessions: When subagents create child sessions, use session_child_first (default: <Leader>+Down) to enter the first child session from the parent.
Once you are in a child session, use:
- session_child_cycle (default: Right) to cycle to the next child session
- session_child_cycle_reverse (default: Left) to cycle to the previous child session
- session_parent (default: Up) to return to the parent session
This lets you switch between the main conversation and specialized subagent work.

Configure

You can customize the built-in agents or create your own through configuration. Agents can be configured in two ways:

JSON

Configure agents in your opencode.json config file:

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "build": {
      "mode": "primary",
      "model": "anthropic/claude-sonnet-4-20250514",
      "prompt": "{file:./prompts/build.txt}",
      "tools": {
        "write": true,
        "edit": true,
        "bash": true
      }
    },
    "plan": {
      "mode": "primary",
      "model": "anthropic/claude-haiku-4-20250514",
      "tools": {
        "write": false,
        "edit": false,
        "bash": false
      }
    },
    "code-reviewer": {
      "description": "Reviews code for best practices and potential issues",
      "mode": "subagent",
      "model": "anthropic/claude-sonnet-4-20250514",
      "prompt": "You are a code reviewer. Focus on security, performance, and maintainability.",
      "tools": {
        "write": false,
        "edit": false
      }
    }
  }
}

Markdown

You can also define agents using markdown files. Place them in:

Global: ~/.config/opencode/agents/
Per-project: .opencode/agents/

---
description: Reviews code for quality and best practices
mode: subagent
model: anthropic/claude-sonnet-4-20250514
temperature: 0.1
tools:
  write: false
  edit: false
  bash: false
---

You are in code review mode. Focus on:

- Code quality and best practices
- Potential bugs and edge cases
- Performance implications
- Security considerations

Provide constructive feedback without making direct changes.

The markdown file name becomes the agent name. For example, review.md creates a review agent.

Environment variables

You can use environment variables in markdown agent frontmatter using the {env:VAR_NAME} syntax. This is useful for dynamic configuration, such as selecting models based on environment.

---
description: "Agent using environment-configured model"
mode: primary
model: "{env:OPENCODE_MODEL}"
---

Your prompt content here.

If the environment variable is not set, it will be replaced with an empty string.

:::note Environment variable interpolation only applies to frontmatter fields, not to the markdown body content. This prevents accidental substitution of {env:VAR} patterns that may appear as examples or documentation in your prompt. :::

This enables workflows like tiered model selection:

# Set different models for different purposes
export OPENCODE_HIGH_MODEL=anthropic/claude-opus-4-5
export OPENCODE_MID_MODEL=anthropic/claude-sonnet-4-20250514
export OPENCODE_LOW_MODEL=anthropic/claude-haiku-4-20250514

Then reference them in your agents:

---
description: "High-capability planning agent"
mode: primary
model: "{env:OPENCODE_HIGH_MODEL}"
---

Options

Let's look at these configuration options in detail.

Description

Use the description option to provide a brief description of what the agent does and when to use it.

{
  "agent": {
    "review": {
      "description": "Reviews code for best practices and potential issues"
    }
  }
}

This is a required config option.

Temperature

Control the randomness and creativity of the LLM's responses with the temperature config.

Lower values make responses more focused and deterministic, while higher values increase creativity and variability.

{
  "agent": {
    "plan": {
      "temperature": 0.1
    },
    "creative": {
      "temperature": 0.8
    }
  }
}

Temperature values typically range from 0.0 to 1.0:

0.0-0.2: Very focused and deterministic responses, ideal for code analysis and planning
0.3-0.5: Balanced responses with some creativity, good for general development tasks
0.6-1.0: More creative and varied responses, useful for brainstorming and exploration

{
  "agent": {
    "analyze": {
      "temperature": 0.1,
      "prompt": "{file:./prompts/analysis.txt}"
    },
    "build": {
      "temperature": 0.3
    },
    "brainstorm": {
      "temperature": 0.7,
      "prompt": "{file:./prompts/creative.txt}"
    }
  }
}

If no temperature is specified, OpenCode uses model-specific defaults; typically 0 for most models, 0.55 for Qwen models.

Max steps

Control the maximum number of agentic iterations an agent can perform before being forced to respond with text only. This allows users who wish to control costs to set a limit on agentic actions.

If this is not set, the agent will continue to iterate until the model chooses to stop or the user interrupts the session.

{
  "agent": {
    "quick-thinker": {
      "description": "Fast reasoning with limited iterations",
      "prompt": "You are a quick thinker. Solve problems with minimal steps.",
      "steps": 5
    }
  }
}

When the limit is reached, the agent receives a special system prompt instructing it to respond with a summarization of its work and recommended remaining tasks.

:::caution The legacy maxSteps field is deprecated. Use steps instead. :::

Disable

Set to true to disable the agent.

{
  "agent": {
    "review": {
      "disable": true
    }
  }
}

Prompt

Specify a custom system prompt file for this agent with the prompt config. The prompt file should contain instructions specific to the agent's purpose.

{
  "agent": {
    "review": {
      "prompt": "{file:./prompts/code-review.txt}"
    }
  }
}

This path is relative to where the config file is located. So this works for both the global OpenCode config and the project specific config.

Model

Use the model config to override the model for this agent. Useful for using different models optimized for different tasks. For example, a faster model for planning, a more capable model for implementation.

:::tip If you don’t specify a model, primary agents use the model globally configured while subagents will use the model of the primary agent that invoked the subagent. :::

{
  "agent": {
    "plan": {
      "model": "anthropic/claude-haiku-4-20250514"
    }
  }
}

The model ID in your OpenCode config uses the format provider/model-id. For example, if you're using OpenCode Zen, you would use opencode/gpt-5.1-codex for GPT 5.1 Codex.

Tools (deprecated)

tools is deprecated. Prefer the agent's permission field for new configs, updates and more fine-grained control.

Allows you to control which tools are available in this agent. You can enable or disable specific tools by setting them to true or false. In an agent's tools config, true is equivalent to {"*": "allow"} permission and false is equivalent to {"*": "deny"} permission.

{
  "$schema": "https://opencode.ai/config.json",
  "tools": {
    "write": true,
    "bash": true
  },
  "agent": {
    "plan": {
      "tools": {
        "write": false,
        "bash": false
      }
    }
  }
}

:::note The agent-specific config overrides the global config. :::

You can also use wildcards in legacy tools entries to control multiple tools at once. For example, to disable all tools from an MCP server:

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "readonly": {
      "tools": {
        "mymcp_*": false,
        "write": false,
        "edit": false
      }
    }
  }
}

Learn more about tools.

Permissions

You can configure permissions to manage what actions an agent can take. Currently, the permissions for the edit, bash, and webfetch tools can be configured to:

"ask" — Prompt for approval before running the tool
"allow" — Allow all operations without approval
"deny" — Disable the tool

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "edit": "deny"
  }
}

You can override these permissions per agent.

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "edit": "deny"
  },
  "agent": {
    "build": {
      "permission": {
        "edit": "ask"
      }
    }
  }
}

You can also set permissions in Markdown agents.

---
description: Code review without edits
mode: subagent
permission:
  edit: deny
  bash:
    "*": ask
    "git diff": allow
    "git log*": allow
    "grep *": allow
  webfetch: deny
---

Only analyze code and suggest changes.

You can set permissions for specific bash commands.

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "build": {
      "permission": {
        "bash": {
          "git push": "ask",
          "grep *": "allow"
        }
      }
    }
  }
}

This can take a glob pattern.

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "build": {
      "permission": {
        "bash": {
          "git *": "ask"
        }
      }
    }
  }
}

And you can also use the * wildcard to manage permissions for all commands. Since the last matching rule takes precedence, put the * wildcard first and specific rules after.

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "build": {
      "permission": {
        "bash": {
          "*": "ask",
          "git status *": "allow"
        }
      }
    }
  }
}

Learn more about permissions.

Mode

Control the agent's mode with the mode config. The mode option is used to determine how the agent can be used.

{
  "agent": {
    "review": {
      "mode": "subagent"
    }
  }
}

The mode option can be set to primary, subagent, or all. If no mode is specified, it defaults to all.

Hidden

Hide a subagent from the @ autocomplete menu with hidden: true. Useful for internal subagents that should only be invoked programmatically by other agents via the Task tool.

{
  "agent": {
    "internal-helper": {
      "mode": "subagent",
      "hidden": true
    }
  }
}

This only affects user visibility in the autocomplete menu. Hidden agents can still be invoked by the model via the Task tool if permissions allow.

:::note Only applies to mode: subagent agents. :::

Task permissions

Control which subagents an agent can invoke via the Task tool with permission.task. Uses glob patterns for flexible matching.

{
  "agent": {
    "orchestrator": {
      "mode": "primary",
      "permission": {
        "task": {
          "*": "deny",
          "orchestrator-*": "allow",
          "code-reviewer": "ask"
        }
      }
    }
  }
}

When set to deny, the subagent is removed from the Task tool description entirely, so the model won't attempt to invoke it.

:::tip Rules are evaluated in order, and the last matching rule wins. In the example above, orchestrator-planner matches both * (deny) and orchestrator-* (allow), but since orchestrator-* comes after *, the result is allow. :::

:::tip Users can always invoke any subagent directly via the @ autocomplete menu, even if the agent's task permissions would deny it. :::

Color

Customize the agent's visual appearance in the UI with the color option. This affects how the agent appears in the interface.

Use a valid hex color (e.g., #FF5733) or theme color: primary, secondary, accent, success, warning, error, info.

{
  "agent": {
    "creative": {
      "color": "#ff6b6b"
    },
    "code-reviewer": {
      "color": "accent"
    }
  }
}

Top P

Control response diversity with the top_p option. Alternative to temperature for controlling randomness.

{
  "agent": {
    "brainstorm": {
      "top_p": 0.9
    }
  }
}

Values range from 0.0 to 1.0. Lower values are more focused, higher values more diverse.

Additional

Any other options you specify in your agent configuration will be passed through directly to the provider as model options. This allows you to use provider-specific features and parameters.

For example, with OpenAI's reasoning models, you can control the reasoning effort:

{
  "agent": {
    "deep-thinker": {
      "description": "Agent that uses high reasoning effort for complex problems",
      "model": "openai/gpt-5",
      "reasoningEffort": "high",
      "textVerbosity": "low"
    }
  }
}

These additional options are model and provider-specific. Check your provider's documentation for available parameters.

:::tip Run opencode models to see a list of the available models. :::

Create agents

You can create new agents using the following command:

opencode agent create

This interactive command will:

Ask where to save the agent; global or project-specific.
Description of what the agent should do.
Generate an appropriate system prompt and identifier.
Let you select which tools the agent can access.
Finally, create a markdown file with the agent configuration.

Use cases

Here are some common use cases for different agents.

Build agent: Full development work with all tools enabled
Plan agent: Analysis and planning without making changes
Review agent: Code review with read-only access plus documentation tools
Debug agent: Focused on investigation with bash and read tools enabled
Docs agent: Documentation writing with file operations but no system commands

Examples

Here are some example agents you might find useful.

:::tip Do you have an agent you'd like to share? Submit a PR. :::

Documentation agent

---
description: Writes and maintains project documentation
mode: subagent
tools:
  bash: false
---

You are a technical writer. Create clear, comprehensive documentation.

Focus on:

- Clear explanations
- Proper structure
- Code examples
- User-friendly language

Security auditor

---
description: Performs security audits and identifies vulnerabilities
mode: subagent
tools:
  write: false
  edit: false
---

You are a security expert. Focus on identifying potential security issues.

Look for:

- Input validation vulnerabilities
- Authentication and authorization flaws
- Data exposure risks
- Dependency vulnerabilities
- Configuration security issues

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Types

Primary agents

Subagents

Built-in

Use build

Use plan

Use general

Use explore

Use compaction

Use title

Use summary

Usage

Configure

JSON

Markdown

Environment variables

Options

Description

Temperature

Max steps

Disable

Prompt

Model

Tools (deprecated)

Permissions

Mode

Hidden

Task permissions

Color

Top P

Additional

Create agents

Use cases

Examples

Documentation agent

Security auditor

FilesExpand file tree

agents.mdx

Latest commit

History

agents.mdx

File metadata and controls

Types

Primary agents

Subagents

Built-in

Use build

Use plan

Use general

Use explore

Use compaction

Use title

Use summary

Usage

Configure

JSON

Markdown

Environment variables

Options

Description

Temperature

Max steps

Disable

Prompt

Model

Tools (deprecated)

Permissions

Mode

Hidden

Task permissions

Color

Top P

Additional

Create agents

Use cases

Examples

Documentation agent

Security auditor