OpenCode — Agent Framework Reference

OpenCode is the AI agent framework running this environment. It manages agents, skills, tools, commands, sessions, providers, and plugins. Everything is configured via files in the config directory (/opt/opencode/ in the Kortix sandbox, or .opencode/ in a project).

Architecture Overview

┌─────────────────────────────────────────────────────┐
│  OpenCode                                            │
│                                                      │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐          │
│  │  Agents   │  │  Skills   │  │  Tools    │          │
│  │ (.md)     │  │ (SKILL.md)│  │ (.ts)     │          │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘          │
│       │              │              │                 │
│  ┌────▼──────────────▼──────────────▼─────┐          │
│  │              Session Engine              │          │
│  │  (prompt → model → tool calls → text)   │          │
│  └────────────────┬───────────────────────┘          │
│                   │                                   │
│  ┌────────────────▼───────────────────────┐          │
│  │            Provider Layer               │          │
│  │  (Anthropic, OpenAI, Kortix router)     │          │
│  └────────────────────────────────────────┘          │
│                                                      │
│  Config: opencode.jsonc  │  Plugins  │  MCP Servers  │
└─────────────────────────────────────────────────────┘

Agents

What is an Agent?

An agent is a persona — a system prompt combined with permission rules, model preferences, and behavioral constraints. Defined as .md files with YAML frontmatter.

Agent Definition File

Location: agents/*.md or agents/**/*.md in the config directory.

---
description: "Short description shown in UI and Task tool"
model: provider/model-id
mode: primary
permission:
  bash: allow
  edit: allow
  read: allow
  task: allow
  skill: allow
  # ... tool permissions
---

# Agent Name

System prompt content goes here. This entire markdown body
becomes the agent's system prompt.

Frontmatter Fields

Agent Modes

Key insight: The Task tool's description filters out primary mode agents (a.mode !== "primary"), but Agent.get() (the execution path) has no mode guard. A primary agent CAN self-spawn by name via subagent_type: "agent-name".

Agent Name Derivation

The agent name is the filename without .md. For nested paths, only the filename matters:

• agents/kortix.md → name: kortix
• agents/specialised/my-agent.md → name: my-agent

Agent Loading Order

1. Built-in agents (compaction, title, summary — hidden infrastructure)
2. Config directory agents (agents/*.md)
3. Config overrides from opencode.jsonc "agent" section (can disable built-in agents)

Disabling Built-in Agents

In opencode.jsonc:

{
  "agent": {
    "build": { "disable": true },
    "plan": { "disable": true },
    "explore": { "disable": true },
    "general": { "disable": true }
  }
}

Agent Permission System

Permissions control which tools an agent can use. Rules are evaluated per-tool-call:

permission:
  bash: allow       # Allow all bash commands
  edit: allow       # Allow file editing
  task: allow       # Allow spawning subagents
  skill: allow      # Allow loading skills
  web-search: allow # Allow web search tool

Subagent sessions inherit parent permissions but can have additional restrictions. The Task tool hardcodes todowrite: deny and todoread: deny for all subagent sessions (lines 77-82 of task.ts).

Skills

What is a Skill?

A skill is a knowledge package — domain-specific instructions, workflows, and resources that inject into the agent's context when loaded via the skill() tool. Skills transform a general-purpose agent into a specialist on demand.

Skill Structure

skill-name/
├── SKILL.md          # Required: frontmatter + instructions
├── scripts/          # Optional: executable code
├── references/       # Optional: supplementary docs
└── assets/           # Optional: templates, images

SKILL.md Format

---
name: my-skill
description: "Detailed description of when to load this skill. Include trigger phrases."
---

# Skill Title

Instructions loaded into context when the skill is triggered.
These become part of the agent's working knowledge.

Frontmatter fields:

Skill Discovery

Skills are discovered from multiple locations (loaded in order, later overwrites earlier):

1. Global external dirs: ~/.claude/skills/**/SKILL.md, ~/.agents/skills/**/SKILL.md
2. Project external dirs: Walk up from project dir looking for .claude/skills/, .agents/skills/
3. OpenCode config dirs: {config}/skills/**/SKILL.md or {config}/skill/**/SKILL.md
4. Additional paths: From opencode.jsonc skills.paths array
5. URL downloads: From opencode.jsonc skills.urls array

Skill Loading Flow

1. On startup, all SKILL.md files are scanned and their name + description are extracted (~100 tokens each)
2. These descriptions are included in the skill tool's description (available_skills list)
3. The agent reads descriptions and decides when to load a skill
4. When loaded via skill({ name: "..." }), the full SKILL.md body is injected into context
5. Bundled files (scripts/, references/) are listed but not loaded — the agent reads them as needed

How the skill tool works

The skill tool (defined in src/tool/skill.ts):

1. Lists all accessible skills in its description (filtered by agent permissions)
2. When called with a skill name, reads the full SKILL.md content
3. Returns <skill_content name="..."> block with the content + skill base directory
4. Also lists up to 10 files in the skill directory for the agent to reference

Tools

What is a Tool?

A tool is a capability the agent can invoke — bash commands, file operations, web search, etc. Tools are defined in TypeScript and registered with the framework.

Built-in Tools

These are part of OpenCode core:

Custom Tools

Custom tools are TypeScript files in tools/*.ts in the config directory. They use the Tool.define() API:

import { Tool } from "opencode/tool"
import z from "zod"

export default Tool.define("my-tool", async (ctx) => {
  return {
    description: "What this tool does",
    parameters: z.object({
      input: z.string().describe("Input parameter"),
    }),
    async execute(params, ctx) {
      // Do work
      return {
        title: "Short result title",
        output: "Full output text",
        metadata: { key: "value" },
      }
    },
  }
})

Tool Permissions

Tools respect the agent's permission rules. Each tool call checks:

1. Agent-level permissions (from agent frontmatter)
2. Session-level permissions (from parent session overrides)
3. Global permission setting (from opencode.jsonc "permission")

With "permission": "allow" in config, all tools are auto-approved.

MCP Tools

OpenCode supports Model Context Protocol (MCP) servers that provide additional tools:

{
  "mcp": {
    "context7": {
      "type": "remote",
      "url": "https://mcp.context7.com/mcp",
      "headers": { "CONTEXT7_API_KEY": "{env:CONTEXT7_API_KEY}" },
      "enabled": true
    }
  }
}

MCP tools appear alongside built-in and custom tools. The agent uses them the same way.

Commands

What is a Command?

A command is a slash-triggered workflow — a prompt template that routes to a specific agent with predefined instructions. Defined as .md files with YAML frontmatter.

Command Definition

Location: commands/*.md in the config directory.

---
description: "What this command does"
agent: kortix
---

# Command Prompt Template

This content becomes the prompt sent to the specified agent.
The user's input after the slash command is appended.

$ARGUMENTS - placeholder for user's input after the command name

Frontmatter Fields

Command Routing

When user types /research quantum computing:

1. OpenCode looks up commands/research.md
2. Reads the frontmatter to get the target agent
3. Sends the command's prompt template (with $ARGUMENTS replaced) to that agent
4. The agent executes with its full tool/skill access

Sessions

Session Lifecycle

CREATE → PROMPT → BUSY (tool calls, text generation) → IDLE → PROMPT again...
                                                      → ABORT
                                                      → DELETE

Session Data Model

{
  "id": "ses_...",
  "title": "Session Title",
  "projectID": "global",
  "directory": "/workspace",
  "parentID": "ses_...",           // Only for subagent sessions
  "permission": [...],             // Permission overrides
  "time": { "created": 123, "updated": 456 }
}

Subagent Sessions

When the Task tool spawns a subagent:

1. Creates a child session with parentID pointing to parent
2. Applies permission restrictions (todowrite/todoread denied, optionally task denied)
3. Sends the prompt to the child session
4. Waits for completion, returns the last text part

Storage Layout

All session data stored as JSON files:

.local/share/opencode/storage/
├── session/global/ses_*.json     # Session metadata
├── message/ses_*/msg_*.json      # Messages per session  
├── part/msg_*/prt_*.json         # Content parts per message
├── todo/ses_*.json               # Todo lists per session
└── tool-output/tool_*            # Large tool outputs

REST API

Base URL: http://localhost:3111

SSE Events

GET /event returns a Server-Sent Events stream:

data: {"type":"session.status","properties":{"sessionID":"ses_...","status":{"type":"busy"}}}
data: {"type":"message.part.updated","properties":{"part":{"type":"text",...},"delta":"word "}}
data: {"type":"session.idle","properties":{"sessionID":"ses_..."}}

Event types: server.connected, session.status, session.idle, session.updated, message.updated, message.part.updated, session.diff, file.edited, command.executed

Configuration

opencode.jsonc

Main config file. In the Kortix sandbox: /opt/opencode/opencode.jsonc

{
  "$schema": "https://opencode.ai/config.json",
  "permission": "allow",                    // Auto-approve all tool calls
  "default_agent": "kortix",               // Default agent for new sessions
  "autoupdate": true,                       // Auto-update OpenCode
  "plugin": [                               // Plugins to load
    "opencode-pty",                         // PTY terminal support
    "./plugin/worktree.ts",                 // Git worktree management
    "./plugin/memory.ts"                    // Memory observation system
  ],
  "agent": {                                // Agent overrides
    "build": { "disable": true },           // Disable built-in agents
    "plan": { "disable": true },
    "explore": { "disable": true },
    "general": { "disable": true }
  },
  "provider": {                             // LLM providers
    "kortix": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "{env:KORTIX_API_URL}/v1/router", "apiKey": "{env:KORTIX_TOKEN}" },
      "models": {
        "kortix/basic": { "name": "Kortix Basic", "cost": { "input": 3, "output": 15 } },
        "kortix/power": { "name": "Kortix Power", "cost": { "input": 5, "output": 25 } }
      }
    }
  },
  "mcp": {                                  // MCP server connections
    "context7": { "type": "remote", "url": "https://mcp.context7.com/mcp" }
  }
}

Config Discovery

OpenCode searches for config in these directories (merged in order):

1. .opencode/ in the current directory and parent directories
2. $OPENCODE_CONFIG_DIR (e.g., /opt/opencode/)
3. Global: ~/.config/opencode/

Environment Variable Interpolation

Config values support {env:VAR_NAME} syntax for environment variable interpolation:

{ "baseURL": "{env:KORTIX_API_URL}/v1/router" }

Plugins

Plugins extend OpenCode with custom tools, hooks, and functionality:

• npm plugins: Referenced by package name (e.g., "opencode-pty")
• local plugins: Referenced by path (e.g., "./plugin/memory.ts")

Plugins can register tools, listen to events, modify behavior.

Provider System

How Providers Work

Providers connect OpenCode to LLM APIs. They use the Vercel AI SDK provider pattern:

{
  "provider": {
    "provider-name": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Display Name",
      "options": { "baseURL": "...", "apiKey": "..." },
      "models": {
        "provider/model-id": {
          "name": "Model Name",
          "cost": { "input": 3, "output": 15 },
          "limit": { "context": 200000, "output": 8192 }
        }
      }
    }
  }
}

Model References

Models are referenced as provider/model-id throughout the system:

• In agent frontmatter: model: anthropic/claude-opus-4-6
• In API calls: { "providerID": "kortix", "modelID": "kortix/basic" }
• In config: "model": "anthropic/claude-sonnet-4-20250514"

Key Files