OpenCode vs Claude Code: 161K Stars, Agent View, and the Provider War (2026)

// Claude Code System Prompt Components
├── Core System Prompt (2,896 tokens)
│   ├── Identity & capabilities
│   ├── Tool usage instructions
│   └── Response formatting rules
│
├── Subagent Prompts
│   ├── Plan Mode Enhanced (633 tokens)
│   ├── Explore Agent (516 tokens)
│   ├── Task Tool (294 tokens)
│   └── Agent View orchestration (new)
│
├── Tool Descriptions (20+ built-in tools)
│   ├── Write, Read, Edit, Bash
│   ├── Task, TodoWrite, Glob, Grep
│   ├── WebFetch, WebSearch
│   ├── NotebookEdit, KillShell
│   ├── EnterWorktree, ExitWorktree
│   └── Plugin-provided tools (dynamic)
│
├── Autonomy Prompts (new in v2.1.139+)
│   ├── /goal evaluator (validator model)
│   ├── Agent View session management
│   ├── /ultrareview code review skill
│   └── /ultraplan cloud environment planner
│
├── Utility Prompts
│   ├── Conversation summarization + recap
│   ├── Session management + persistence
│   ├── CLAUDE.md creation (384 tokens)
│   ├── Security review (2,610 tokens)
│   └── Agent creation architect (1,110 tokens)
│
└── Slash Command Prompts (expanded)
    ├── /goal, /loop, /focus, /effort
    ├── /ultrareview, /ultraplan
    ├── /pr-comments, /review-pr
    ├── /team-onboarding
    └── /security-review (2,610 tokens)

# .opencode/agents/build.md (Primary Agent)
---
name: build
description: Default agent for development work
mode: primary
model: anthropic/claude-sonnet-4  # Or any provider
temperature: 0.7
tools:
  write: true
  edit: true
  bash: true
  webfetch: true
  task: true
  todo: true
permission:
  edit: allow
  bash: ask
  webfetch: allow
maxSteps: 50
---

You are a coding assistant focused on software development.
[Custom instructions follow...]

# .opencode/agents/plan.md (Read-only Agent)
---
name: plan
description: Analysis and exploration agent
mode: primary
permission:
  edit: deny
  bash: ask
---

You are an analyst. Read code, explain patterns,
but do not modify files unless explicitly asked.

// Claude Code v2.1.143 Built-in Tools
const CLAUDE_CODE_TOOLS = [
  // File Operations
  "Read",           // Read files with line ranges
  "Write",          // Create/overwrite files
  "Edit",           // Search-replace edits
  "Glob",           // Find files by pattern (native binary)
  "Grep",           // Search file contents (native binary)

  // Execution
  "Bash",           // Execute shell commands
  "KillShell",      // Terminate running shells
  "NotebookEdit",   // Edit Jupyter notebooks

  // Agent System
  "Task",           // Spawn subagents
  "TodoWrite",      // Track task progress

  // Worktree Management (new)
  "EnterWorktree",  // Isolate background work
  "ExitWorktree",   // Return to main working copy

  // Web & External
  "WebFetch",       // Fetch web content
  "WebSearch",      // Search the web

  // IDE Integration
  "AskUserQuestion", // Request user input
  "EnterPlanMode",   // Switch to planning
  "ExitPlanMode",    // Exit planning

  // Plugin Tools (new in 2.1.117+)
  "plugin__*",      // Plugin-provided tools/skills

  // MCP Tools
  "mcp__*",         // Dynamically loaded MCP tools
];

// OpenCode v1.15.0 Built-in Tools
const OPENCODE_TOOLS = [
  // File Operations
  "read",           // Read files with line ranges
  "write",          // Create/overwrite files
  "edit",           // Search-replace edits
  "glob",           // Find files by pattern
  "grep",           // Search with ripgrep

  // Execution
  "bash",           // Execute shell commands

  // Agent System
  "task",           // Spawn subagents (General, Explore, Scout)
  "todo",           // Track task progress
  "skill",          // Load reusable instruction sets

  // Web
  "webfetch",       // Fetch and process web content

  // MCP Tools
  "mcp_*",          // Declaratively configured MCP tools
];

// Built-in Subagents (v1.14.42+):
// General  - full tool access, multi-step tasks
// Explore  - read-only codebase navigation
// Scout    - read-only external docs & dependency research

// Claude Code: MCP tools via mcp.json
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"]
    }
  }
}
// All tools loaded into context at startup
// New: Tool Search enables lazy loading

// OpenCode: Declarative per-agent control
// opencode.json
{
  "tools": {
    "mymcp_*": true,          // Enable for all agents
    "dangerous_tool": false   // Disable globally
  },
  "agent": {
    "build": {
      "tools": {
        "mymcp_write": true,
        "mymcp_delete": false  // Disable for this agent
      }
    }
  }
}

// Claude Code Subagent Architecture (May 2026)
Agent View Dashboard (new: v2.1.139)
│   ├── Shows all sessions: Running / Blocked / Done
│   ├── Navigate between sessions
│   └── Peek at results without switching
│
├── /goal Command (new: v2.1.139)
│   ├── Set completion condition
│   ├── Validator model checks after every step
│   ├── Claude works autonomously until goal met
│   └── Enables true fire-and-forget workflows
│
├── Plan Mode (633 tokens prompt)
│   ├── Triggered via /plan or EnterPlanMode
│   ├── Read-only by default
│   ├── /ultraplan for cloud environment planning
│   └── Exits with ExitPlanMode tool
│
├── Explore Agent (516 tokens prompt)
│   ├── Fast codebase navigation
│   ├── Read-only access
│   └── Returns findings to parent
│
├── Task Agent (294 tokens prompt)
│   ├── Complex multi-step workflows
│   ├── Own context window + worktree isolation
│   ├── Background sessions with --bg flag
│   └── Results summarized to parent
│
├── Custom Agents (user-defined)
│   ├── Defined in .claude/agents/
│   ├── Inherit tool access from config
│   └── Agent-specific hooks and MCP servers
│
└── Plugin Agents (new: v2.1.117+)
    ├── Installed from plugin marketplace
    ├── Provide skills, hooks, and MCP servers
    └── Dependency enforcement between plugins

// Key Features:
// Esc twice = instant rewind to last checkpoint
// /loop = proactive monitoring mode
// /focus = distraction-free fullscreen
// Worktree isolation for parallel sessions

// OpenCode Agent Architecture (May 2026)
// KEY ADVANTAGE: Superior interactivity + Scout

Primary Agents (Tab to switch instantly)
├── Build Agent (default)
│   ├── All tools enabled
│   ├── Full filesystem access
│   └── For development work
│
└── Plan Agent
    ├── Read-only by default
    ├── Asks permission for bash
    └── For analysis/exploration

Subagents (via Task tool or @mention)
├── @general - Complex multi-step tasks (parallel)
├── @explore - Read-only codebase navigation
├── @scout  - External docs & dependency research (new)
│   ├── Clones dependency repos into cache
│   ├── Inspects library source code
│   └── Read-only, never modifies workspace
└── @custom - Your own agents from .opencode/agents/

// Background Subagents (new: v1.14.51)
// Tasks keep running while you continue working
// Session pinning with quick slots (v1.14.49)

// INTERACTIVITY FEATURES:
// 1. Tab - instant agent switching
// 2. @agent_name - invoke any subagent inline
// 3. HTTP API - control from mobile/remote
// 4. Persistent sessions - survives terminal close
// 5. Session pinning - quick slots for fast access
// 6. Auto-compact - intelligent context summarization

// Claude Code uses automatic compaction
// Triggered near context limit

// Compaction process:
1. Detect approaching context limit
2. Summarize previous messages (utility prompt)
3. Preserve critical context (file contents, etc.)
4. Continue with reduced token count

// Manual compaction:
$ /compact     # Slash command to force compaction

// Context advice from Anthropic:
// "If using Claude in an agent harness that compacts context,
// add this to your prompt so Claude can behave accordingly.
// Otherwise Claude may try to wrap up work as it
// approaches the context limit."

// OpenCode now has auto-compact (v1.14+)
// Plus provider-dependent Vercel AI SDK handling

// Auto-compact:
// - Intelligently summarizes conversation history
// - Manages token usage across long sessions
// - System agent handles compaction automatically

// Session persistence:
// - Server-side state via Hono HTTP API
// - Sessions survive terminal closure
// - Session pinning + quick slots (v1.14.49)
// - Can reconnect from mobile app

// LSP Integration:
// - Automatic symbol awareness
// - Language-specific context enrichment
// - Better code navigation than text-based approaches

// claude.permissions in settings.json
{
  "claude.permissions": {
    "edit": "ask",           // ask | allow | deny
    "bash": "ask",
    "bash.rm": "deny",       // Specific command patterns
    "bash.git": "allow",
    "webfetch": "allow"
  }
}

// Permission flow:
1. Tool call requested
2. Check permission config
3. If "ask": Show confirmation dialog
4. If "allow": Execute immediately
5. If "deny": Block with error

// Dangerous operation detection:
// Built-in heuristics for destructive commands
// Even with "allow", warns on rm -rf /, etc.

// opencode.json permission configuration
{
  "permission": {
    "edit": "allow",
    "bash": {
      "*": "ask",              // Default: ask for all
      "git *": "allow",        // Allow git commands
      "npm test": "allow",
      "npm install": "ask",
      "rm -rf *": "deny",      // Block dangerous patterns
      "sudo *": "deny"
    },
    "webfetch": "allow",
    "task": {
      "general": "allow",
      "explore": "allow",
      "custom-*": "ask"
    },
    "skill": {
      "pr-review": "allow",
      "dangerous-*": "deny"
    }
  }
}

// Permission inheritance:
// Agent-specific overrides in frontmatter
// Last matching glob pattern wins

// Claude Code: mcp.json
{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": { "DATABASE_URL": "..." }
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"]
    }
  }
}

// OpenCode: opencode.json
{
  "mcp": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": { "DATABASE_URL": "..." }
    }
  },
  "agent": {
    "build": {
      "tools": {
        "postgres_query": true,
        "postgres_execute": false  // Read-only for this agent
      }
    }
  }
}