[Read the docs →](https://cli.devin.ai/docs) # Devin CLI The AI software engineer in your terminal, with every frontier model. [READ THE DOCS →](https://cli.devin.ai/docs) New · Fusion in Devin CLI ## Frontier results for a third less. A frontier model makes the decisions while a cheaper sidekick does the legwork, so you pay frontier prices only where they matter. [How Fusion works →](https://cognition.ai/blog/fusion-in-devin-cli)[Artificial Analysis results →](https://artificialanalysis.ai/)[FrontierCode benchmarks →](https://cognition.ai/frontiercode) 36%Lower cost than Claude Code with Fable 5.1. 39%Lower cost than Codex with Astra 6. Independently measured by [Artificial Analysis](https://artificialanalysis.ai/) on the Coding Agent Index 1.5. Cost per run · [AA Coding Agent Index 1.5](https://artificialanalysis.ai/)Score - - Claude Code · Fable 5.1 - $12.3662.2 - Devin Fusion · Fable 5.1 + SWE-2 - $7.9061.7 * * Codex · Astra 6 * $7.4761.6 * Devin Fusion · Astra 6 + SWE-2 * $4.5458.9 On the Artificial Analysis Coding Agent Index 1.5, Devin Fusion with Fable 5.1 cost $7.90 per run versus $12.36 for Claude Code with Fable 5.1, and $4.54 with Astra 6 versus $7.47 for Codex with Astra 6, at frontier scores. Multi-model ## Every frontier model in your CLI Anthropic, OpenAI, Google, Cognition, and open-weight models, switchable mid-session with /model. devin — \~/acme/web ❭ /model / Type to search ↑ more above ·Claude Opus 5 ◼◼◼◼ Medium Fast ModeOff ·Claude Sonnet 5 ◼◼◼◼ Medium ·DeepSeek V4 Flash ◼◼◼ High ·DeepSeek V4 Pro ◼◼◼ High ❭Fusion✱←◼◼◼◼→HighLeadClaude Fable…▾SidekickSWE-2 High▾ Fast ModeOff ·Gemini 3.7 Flash ◼◼◼ Medium ·GLM-5.3 ◼◼◼ Max ·GPT-5.6 Luna ◼◼◼◼◼ Medium Fast ModeOff ·GPT-5.6 Sol ◼◼◼◼◼ Medium Fast ModeOff ↓ more below LeadClaude Fable 5.1▾SidekickSWE-2 High▾ Fast ModeOff InputCached inputOutput $10 / 1M$0.25 / 1M$50 / 1M ✱ New✱ Promotion✱ Beta ↑↓ select · tab lead · ←→ effort · ↵ confirm · esc cancel Devin, locally ## The intelligence of Devin, in the terminal you already use. ### Runs on your machine Your repo, your shell, your credentials. ### Hand off to the cloud Close your laptop, come back to a PR. ### Built in Rust Installs in under a minute, starts instantly. ### Bring your setup MCP servers, skills, hooks, and Claude Code plugins work as-is. Everything in the box ## One install. The whole toolkit. Every command, mode, model, and integration Devin CLI ships with. Commands - /modelAny model, effort level, and Fusion pairing - /handoffContinue the task in Devin Cloud - /loopImplement, then auto-review the diff until it's right - /btwAsk a side question without derailing the task - /planAgree on the approach before any edits - /askRead-only mode for exploring a codebase - /forkBranch the session from any step - /revertRoll files and conversation back to a step - /recapCatch up on a long session in a paragraph - /resumePick up any previous session - /compactCompress context without losing the thread - /contextSee exactly what is in the window - /shareSend the session to your team on Devin - /mcpCheck every connected MCP server - /fastJump to the fastest model you have - devin -pSingle-turn mode for scripts and CI Control - NormalReads auto-approve, writes and shell ask first - Accept EditsAuto-approve edits inside the workspace - Smart modeA fast model judges what is safe to run - BypassFull autonomy when you want it - AutonomousAuto-run inside an OS-level sandbox - \--sandboxFail-closed isolation for files and network - SubagentsParallel workers, foreground or background - Custom subagentsYour own profiles, prompts, and models - Permission rulesAllow, ask, or deny per tool and path - Team settingsAdmin-enforced rules across the org - HooksShell or LLM checks at every lifecycle step - @ filesAttach files and folders as context - Image pasteCtrl+V a screenshot into the prompt - Background shellsLong commands keep running while Devin works Works with - AnthropicClaude Fable 5.1, Opus 5, Sonnet 5 - OpenAIGPT-6 Astra, GPT-5.6 Sol and Luna - GoogleGemini 3.7 - CognitionSWE-2, Fusion, Adaptive - Open weightsKimi, GLM, DeepSeek - AGENTS.mdRules load from the repo automatically - Claude Code pluginsSkills, hooks, and subagents load as-is - Agent Plugins 1.0Install and share bundles across projects - MCP serversstdio and Streamable HTTP - SkillsReusable workflows as slash commands - Zed · JetBrains · XcodeRuns as an ACP agent inside your editor - Devin DesktopSame agent, IDE surface - Devin CloudHand off and track cloud agents from the tray - macOS · Linux · Windowscurl, Homebrew, or PowerShell get started ## Install Devin CLI in under a minute [Read Docs](https://cli.devin.ai/docs) ## Frequently Asked Questions What is Devin CLI? Devin running locally in your terminal: your repo, your shell, your tools. Hand a session off to Devin Cloud when you want it to keep going without your machine. [Read the docs →](https://cli.devin.ai/docs) What is Devin Fusion? A multi-model harness in Devin CLI. A frontier lead model makes the decisions and the important edits; a cheaper sidekick handles exploration, file reads, and test runs in its own context. [Read about how Fusion works →](https://cognition.ai/blog/fusion-in-devin-cli) How does the hand-off to Devin Cloud work? Run /handoff in any session. It moves to Devin Cloud with the same context and keeps working; pick it up later in the web app or Devin Desktop. How is Devin CLI different from Claude Code or Codex? Each is tied to one lab's models. Devin CLI runs Anthropic, OpenAI, Google, Cognition, and open-weight models, combines them with Fusion, and can hand a session off to Devin Cloud. Can I reuse my existing setup? Yes. AGENTS.md rules, MCP servers, and Claude Code plugins (skills, hooks, subagent profiles) load as-is. [Extensibility docs →](https://cli.devin.ai/docs/extensibility) Which platforms are supported? macOS, Linux, and Windows. Install with the one-line script above. [Installation guide →](https://cli.devin.ai/docs)