Droid vs Claude Code: which terminal agent should you run?

CLI comparison at a glance

The commands look similar because both products implement the same agent pattern: receive a goal, inspect context, call tools, observe results, and continue until done. Differences appear in the model relationship, policy system, remote environment, orchestration, and ecosystem. Those layers influence real completion more than the color of the terminal UI.

Installation and authentication

Droid offers a standalone shell installer for macOS and Linux, a PowerShell installer for Windows, a Homebrew cask, and a global npm package. Claude Code offers Anthropic’s native installer and package paths documented for its supported systems. Both open a browser sign-in for subscription-backed interactive use and support automation credentials through their respective platforms.

# Droid on macOS or Linux
curl -fsSL https://app.factory.ai/cli | sh

# Droid through npm
npm install -g droid

# Claude Code native installer on macOS or Linux
curl -fsSL https://claude.ai/install.sh | bash

Droid’s npm build disables automatic updates, which is useful when an exact version is installed. Standalone Droid installations update automatically by default and can be pinned or rolled back with droid update --version. Claude Code also has update management and managed deployment guidance. In CI, pin whichever installation artifact your organization has tested.

Droid unattended work can use FACTORY_API_KEY. Claude automation can use supported Anthropic API, cloud, or organization authentication paths. Keep credentials in a secret manager and give the job the minimum repository, provider, and external-tool access. Authentication success says nothing about authorization safety.

Interactive terminal experience

Droid opens a full-screen TUI with interaction mode, autonomy, model, MCP status, prompt composer, diff approval, transcript, and Mission controls. Slash commands manage review, settings, model, sessions, forks, compression, Missions, custom Droids, skills, hooks, plugins, MCP, account, billing, and limits. Direct shell mode runs a command without model interpretation.

Claude Code opens an interactive terminal session with conversation, tool proposals, permission prompts, plan mode, model controls, context status, tasks, skills, hooks, plugins, MCP, session management, and worktree support. The product also integrates directly into VS Code and JetBrains while preserving the same underlying Claude Code session concepts.

Droid’s interface foregrounds Factory concepts such as autonomy, custom Droids, Missions, and billing. Claude Code foregrounds Claude concepts such as permission modes, Claude models, tasks, plugins, and the Anthropic cloud connection. Both are learnable within one real ticket. Preference is secondary to whether tool use stays legible during a failure.

Model choice and harness quality

Droid is model independent. The model flag selects a supported ID, and enterprise configurations can use direct providers, Bedrock, Vertex, Azure OpenAI, gateways, approved custom models, or on-premises endpoints. Factory can tune prompting and tools for different models. A separate model and reasoning effort can be selected for Specification Mode, and Missions can select worker and validator models independently.

Claude Code uses Claude models through Anthropic’s first-party agent. This reduces ambiguity about which vendor owns model behavior, agent prompting, cloud tasks, and support. Claude Code can switch among Claude models allowed by the plan and organization. It does not serve as a general harness for GPT, Gemini, or an arbitrary enterprise model endpoint.

Factory’s historical Terminal-Bench report showed that Droid could outperform other harnesses using the same model under the disclosed setup. The result supports the importance of prompts, tools, timeouts, environment discovery, planning, and context. It does not establish a permanent winner. Claude Code and Claude models have continued to change, as have Droid and the benchmark.

• Pick Droid when switching model providers without switching terminal workflow is a requirement.
• Pick Claude Code when direct Claude behavior, feature delivery, and support are the requirement.
• Benchmark both on private tasks when the model is the same. This isolates some of the harness difference.
• Re-test after upgrades because agent prompts and tool behavior can shift without repository changes.

Specification Mode versus Plan Mode

Droid Specification Mode asks for a detailed specification before code changes. It can produce acceptance criteria, implementation detail, a file-by-file plan, tests, and security considerations. The user can refine the proposal and approve into an autonomy level. Plans can be saved as Markdown under a Factory docs directory.

Claude Code Plan Mode explores without source edits and presents a plan for approval. The user can continue planning, approve into manual or edit-accepting execution, or use eligible auto and browser planning paths. Plans can be edited in the user’s editor and can seed a cloud execution workflow.

Factory wins when a formal specification artifact and distinct planning model are central. Claude wins when planning should remain one fluid stage in the Claude session and hand off to Claude web execution. For both, inspect the plan for facts that are cheap to verify before authorizing expensive work.

Permissions and unattended work

Droid uses Off, Low, Medium, and High autonomy. Headless execution defaults read-only. Low covers safe edits, Medium covers dependencies, builds, tests, and local commits, and High covers push, deployment, long operations, subagents, and Missions. Tool restrictions, hooks, deny lists, MCP policy, organization caps, and sandboxes layer on top.

Claude Code has default, acceptEdits, plan, auto, dontAsk, and bypassPermissions modes. Allow, ask, and deny rules can target tools and command patterns. Managed settings can cap dangerous behavior. Auto uses a safety classifier for eligible accounts; bypass skips the permission layer and is intended for isolated environments.

Neither permission label is an OS security boundary by itself. A worktree isolates repository files from another branch and shares the host account, network, processes, caches, and credentials. Use a dedicated container, VM, runner, or hardened computer for unattended commands with broad authority.

Headless execution and machine output

Droid Exec treats the agent as a command. It accepts direct prompts, prompt files, stdin, a working directory, model and reasoning controls, tool restrictions, session continuation, forks, tags, worktrees, Specification Mode, and Missions. It returns text, one JSON result, streaming events, or a bidirectional JSON-RPC protocol.

Claude Code print mode uses claude -p for scripts and CI. It supports JSON and streaming output, tool and permission controls, session continuation, model settings, system-prompt additions, turn and budget controls, and schema-constrained structured output. The Claude Agent SDK provides a higher-level programmatic route.

# Review as structured output
droid exec -o json "review the current diff"

claude -p "review the current diff" --output-format json

# Edit in a worktree
droid exec --worktree fix-412 --auto medium "fix issue 412 and run tests"

claude --worktree fix-412 -p "fix issue 412 and run tests" --permission-mode acceptEdits

Droid has the edge for applications that want to drive a long-lived agent process over raw JSON-RPC and answer permission requests in a custom interface. Claude has the edge for schema-constrained result generation and Anthropic SDK alignment. Both need timeouts, process cleanup, version pinning, deterministic post-checks, and sanitized logs.

Sessions, forks, and worktrees

Droid can resume the latest or a named session, continue an exec session by ID, fork a session into a new copy, search local session content, tag runs, and group logs. Claude Code can resume or continue local sessions, fork conversations through its session controls, name sessions, and move cloud work into a terminal through teleport.

Both CLIs now expose --worktree. A named worktree gives one session its own checkout and branch while sharing repository history. Claude Code also supports worktree-isolated subagents and Desktop sessions. Droid can use worktrees for interactive and exec runs, including concurrent evaluation branches.

1. Start clean. Log repository root, branch, status, and worktree list before the agent runs.
2. One writer per worktree. Subagents that edit should receive separate worktrees or explicit non-overlapping ownership.
3. Separate ports and services. The checkout is isolated; test databases and localhost ports are shared unless configured otherwise.
4. Review before cleanup. Preserve the diff, transcript, checks, and failure evidence until the branch is accepted or rejected.

Droid’s session search and tagging are useful in a Factory fleet. Claude’s teleport is useful when a task begins in an Anthropic cloud environment and needs local continuation. The better lifecycle depends on where work begins and where a developer wants to intervene.

Subagents and multi-agent work

Droid supports custom Droids as specialized subagents and Factory Missions as orchestrated multi-agent projects. A Mission has workers, validators, milestones, Mission Control, and headless flags. Worker and validator model settings can differ. High autonomy is required for orchestration under current Factory documentation.

Claude Code supports custom subagents with their own descriptions, prompts, tools, and contexts, plus agent teams and parallel worktree or cloud session patterns. It is effective for delegating exploration, review, and bounded implementation from a parent Claude session. Anthropic also supports scheduled cloud tasks and parallel remote runs.

Parallelism makes review capacity the constraint. Keep tasks narrow, make workers own distinct files or contracts, and require executable checks. A validator model agreeing with a worker model is supporting evidence, not acceptance.

Remote and cloud continuation

Factory can target persistent Droid Computers. Managed computers are provisioned by Factory; BYOM registers an owned Linux, macOS, or Windows machine. Files, packages, services, and configuration persist across sessions. CLI commands list computers, connect over SSH, and forward TCP ports through the relay.

Claude Code can create managed web sessions with claude --remote, monitor them from claude.ai or mobile, and pull a cloud session into the terminal with --teleport. Remote Control exposes a local CLI session for steering through web and mobile while the process remains on the local host. Cloud sessions can continue after the laptop closes and can create pull requests.

• Droid advantage: explicit persistent computer abstraction and customer-owned machine registration.
• Claude advantage: direct terminal-to-claude.ai task creation, mobile monitoring, and teleport into local CLI.
• Droid risk: persistent environment drift and host operations.
• Claude risk: cloud resource, repository, network, and session-lifecycle constraints documented for hosted tasks.

Choose persistence when recreating the environment is the dominant cost. Choose managed task environments when clean setup and provider-owned operations matter. Measure time to first useful tool call, reproducibility, idle behavior, credential handling, and recovery from a failed environment.

Instructions, MCP, hooks, and plugins

Droid reads AGENTS.md and Factory settings, and can load skills, custom Droids, plugins, hooks, MCP servers, and integrated organization context. Claude Code reads CLAUDE.md and settings, and supports skills, plugins, hooks, subagents, MCP, auto memory, and IDE context. Both can encode repository conventions and connect external services.

AGENTS.md has value when instructions should travel among several coding agents. CLAUDE.md is optimized for the Claude Code ecosystem. A team can keep shared build and test facts in AGENTS.md and Claude-specific workflow detail in CLAUDE.md, while avoiding duplicated contradictory rules.

Droid plugins and custom Droids align with Factory’s cross-model harness. Claude plugins and subagents align with the first-party Claude behavior. Prefer repository-owned plain text for durable conventions and use product-specific extensions for procedures that genuinely depend on that runtime.

Pricing, limits, and value

Factory Pro costs $20 per month, Plus $100, and Max $200. Claude Pro costs $20 month to month, and Claude Max tiers cost $100 and $200. The matching dollar ladder makes sticker price a poor differentiator. Existing subscription ownership matters: a Claude subscriber may already have Claude Code access, while a Factory subscription buys Droid and its multi-model platform.

Factory Standard Usage uses rolling 5-hour, weekly, and monthly windows. Droid Core provides a separate included fallback pool after Standard Usage, and prepaid Extra Usage continues supported models. Claude subscription usage has rolling limits shared with Claude surfaces, with other paid access paths available under current plan rules.

Compare a week of accepted work. Count limit hits, fallback quality, overflow spend, remote compute, corrections, reviewer time, and defects. A cheaper session that produces a rejected diff has negative value. An expensive session that safely completes a high-cost migration may be trivial relative to the outcome.

Winner by scenario

Most individual developers should begin with the subscription they already own. If neither exists, both start at $20, so choose the architecture that fits the next year: Droid for model and platform independence, Claude Code for the Anthropic stack. Give the other tool a controlled trial when a concrete workflow remains weak.

Questions people ask

Keep reading