Lesson 3: dcode vs Claude Code

Understanding the similarities, differences, and when to use each

Both dcode and Claude Code are terminal-based coding agents with similar architecture. The key differences: dcode is model-agnostic, Claude Code is Claude-locked; dcode has first-class remote sandboxing, Claude Code executes locally.

Side-by-Side Architecture

Architectural Comparison

flowchart TB subgraph DCODE["dcode (DeepAgents Code)"] D_USER["~/.deepagents/"] D_AGENTS["AGENTS.md"] D_SKILLS["skills/"] D_STATE["sessions.db (SQLite)"] D_LOOP["Agent Loop"] D_SANDBOX["Remote Sandbox
(7+ providers)"] D_MODEL["Any LLM
provider:model"] end subgraph CLAUDE["Claude Code"] C_USER["~/.claude/"] C_AGENTS["CLAUDE.md"] C_SKILLS2["skills/"] C_STATE2["JSON state"] C_LOOP2["Agent Loop"] C_LOCAL["Local Execution"] C_MODEL2["Claude Models
(Opus, Sonnet, Haiku)"] end D_USER --> D_AGENTS D_USER --> D_SKILLS D_USER --> D_STATE D_LOOP --> D_SANDBOX D_LOOP --> D_MODEL C_USER --> C_AGENTS C_USER --> C_SKILLS2 C_USER --> C_STATE2 C_LOOP2 --> C_LOCAL C_LOOP2 --> C_MODEL2 style DCODE fill:#e8f5e9,stroke:#76b900,stroke-width:2px style CLAUDE fill:#e3f2fd,stroke:#1a5fb4,stroke-width:2px

Feature Comparison Matrix

Similarities (What They Share)

Feature dcode Claude Code
Form Factor Terminal CLI Terminal CLI
Instructions File AGENTS.md CLAUDE.md
Skills Framework ~/.deepagents/skills/ ~/.claude/skills/
Human-in-the-Loop Approval controls Permission system
MCP Integration Full support Full support
Context Management Compaction Summarization
Subagent Delegation Supported TaskCreate
Persistent Memory Cross-session Cross-session

Differences (What Sets Them Apart)

Aspect dcode Claude Code
Model Support Any LLM (provider:model format) Claude only (Opus, Sonnet, Haiku)
Sandbox Support 7+ providers (LangSmith, Modal, E2B...) Local + optional worktree
Tracing/Observability Native LangSmith None built-in
Programmatic API create_deep_agent() SDK CLI only
Session State SQLite database JSON files
Harness Tuning Per-model optimization Fixed for Claude
Governed Runtime OpenShell kernel-level Local permissions
Cost at Scale 10x cheaper (with Nemotron) Claude API pricing

The Sandbox-as-Tool Pattern

This is the key architectural difference that enables remote sandboxing in dcode:

dcode: Sandbox-as-Tool Pattern

flowchart LR subgraph LOCAL["Local Machine"] LOOP["dcode Process
(LLM loop, memory, dispatch)"] end subgraph REMOTE["Remote Sandbox"] FS["Filesystem
read_file, write_file"] EXEC["Execution
shell commands"] TESTS["Testing
run tests"] end LOOP -->|"Tool calls"| FS LOOP -->|"Tool calls"| EXEC LOOP -->|"Tool calls"| TESTS style LOCAL fill:#e8f5e9,stroke:#76b900,stroke-width:2px style REMOTE fill:#e3f2fd,stroke:#1a5fb4,stroke-width:2px

"Deep Agents Code uses the 'sandbox as tool' pattern: the dcode process (LLM loop, memory, tool dispatch) runs locally, but agent tool calls (read_file, write_file, execute, etc.) target the remote sandbox, not the local filesystem." — LangChain Remote Sandboxes

Claude Code: Local Execution Pattern

flowchart LR subgraph LOCAL2["Local Machine"] LOOP2["Claude Code Process
(LLM loop, memory, dispatch)"] FS2["Local Filesystem
read_file, write_file"] EXEC2["Local Execution
shell commands"] TESTS2["Local Testing
run tests"] end LOOP2 -->|"Tool calls"| FS2 LOOP2 -->|"Tool calls"| EXEC2 LOOP2 -->|"Tool calls"| TESTS2 style LOCAL2 fill:#e3f2fd,stroke:#1a5fb4,stroke-width:2px

Configuration Comparison

dcode Configuration

# ~/.deepagents/config.toml
default_model = "nvidia:nemotron-3-ultra"
sandbox_provider = "langsmith"

# Resolution order:
# 1. DEEPAGENTS_CODE_* env var (highest)
# 2. Canonical env var
# 3. ~/.deepagents/config.toml
# 4. Built-in default (lowest)

Claude Code Configuration

// ~/.claude/settings.json
{
  "model": "claude-sonnet-4-20250514",
  "permissions": {
    "allow": ["Bash(git *)"],
    "deny": []
  }
}

// Resolution order:
// 1. settings.local.json (highest)
// 2. settings.json
// 3. Built-in defaults (lowest)

Directory Structure Comparison

dcode Directory Structure

~/.deepagents/
├── .state/
│   ├── sessions.db          # SQLite for checkpoints
│   └── history.jsonl        # Command history
└── {agent}/
    ├── AGENTS.md            # User instructions
    ├── skills/              # User-level skills
    └── agents/              # Subagent definitions

{project}/
├── AGENTS.md                # Project instructions (root)
└── .deepagents/
    ├── AGENTS.md            # Project instructions (preferred)
    ├── skills/              # Project skills
    └── agents/              # Project subagents

Claude Code Directory Structure

~/.claude/
├── settings.json            # User settings
├── keybindings.json         # Key customizations
├── CLAUDE.md                # Global instructions
├── skills/                  # User skills
└── projects/                # Project state

{project}/.claude/
├── settings.json            # Project settings
├── settings.local.json      # Local overrides (gitignored)
└── CLAUDE.md                # Project instructions

When to Choose Which

Choose dcode / NemoClaw When:

Choose Claude Code When:

The NemoClaw Advantage

When you combine dcode with NemoClaw's OpenShell runtime, you get capabilities neither tool has alone:

NemoClaw Stack Advantages

flowchart TB subgraph NEMOCLAW["NemoClaw + dcode Stack"] ADV1["Open Model Ownership
Tune, run, optimize Nemotron"] ADV2["Full Stack Control
Tools, context, evals, runtime, policies"] ADV3["10x Cost Efficiency
$4.48 vs $43.48 benchmark"] ADV4["Kernel-Level Security
Landlock LSM, deny-by-default"] ADV5["Multi-Cloud Portability
Cloud, on-prem, RTX, DGX"] end style NEMOCLAW fill:#e8f5e9,stroke:#76b900,stroke-width:2px

"Teams need full control over tools, context, evaluation methods, runtime location, and action policies - closed ecosystems prevent this. Companies need to own that work, improve it over time, and run agents with the controls their organizations require." — LangChain Blog

Benchmark Comparison

Configuration Aggregate Score Cost Cost Efficiency
Nemotron 3 Ultra + dcode 0.86 $4.48 Baseline
Next closest alternative ~0.86 $43.48 ~10x more expensive

Source: LangChain NemoClaw Benchmark

Knowledge Check

Primary Source Reading

Recommended Reading

LangChain DeepAgents Code Overview

https://docs.langchain.com/oss/python/deepagents/code/overview

This page covers dcode's architecture, sandbox providers, and configuration in detail.

Related Reference