Devin

Devin is a hosted agent platform. A cloud session runs in its own VM with a shell, browser and repository access, and has an embedded IDE where a person can follow or take over. Sessions are the unit of work: they can be started from the web app, Slack or Teams, Linear or Jira tickets, automations, the API, MCP clients, or the local Devin CLI.

Two mechanisms give multi-agent behaviour. With managed Devins, any session can act as coordinator and launch child sessions with a prompt, playbook, tags and an ACU (compute) limit. The coordinator can message, sleep or terminate children and schedule reminders for itself. With Dynamic Workflows the coordination is moved into a Python script that Devin writes and runs: agents are called in fan-out or staged patterns, results are structured, and each call is recorded so the run is observable and resumable. The docs recommend a workflow when roughly five or more independent units need judgement and a combine step, and a plain session otherwise.

State and memory sit in the platform: session history, playbooks (reusable procedures), a knowledge base, secrets and organization settings. The local Devin CLI is a separate local agent with foreground and background subagents; it shares Devin accounts and can hand work to cloud sessions.

Managed Devins launch child sessions without asking by default; turning off Auto-approve child sessions in Settings, Preferences makes Devin ask before each batch. A person can message any running session, and the coordinator can pause or terminate children. Devin's workspace has an embedded IDE where you can take over from the agent. In Devin CLI, foreground subagents let you approve or deny tool calls, background subagents automatically deny unapproved tools, and permission rules can allow, deny or prompt. Enterprise accounts must enable Dynamic Workflows before they run.

curl -fsSL https://cli.devin.ai/install.sh | bash

# Devin CLI (macOS, Linux, WSL); Homebrew: brew install --cask devin-cli
curl -fsSL https://cli.devin.ai/install.sh | bash

# In a project directory, start it with a first prompt
devin -- check out this code and suggest a feasible, helpful feature

# Inside the CLI, hand a longer task to a cloud Devin session
# /handoff fix the flaky integration tests in CI

# Cloud Devin: sign up at https://app.devin.ai, then ask a session to parallelize:
# "Start a parallel Devin session for each module below 50% test coverage"

• The cloud product is used from https://app.devin.ai; the CLI is a separate local agent that needs a Devin login.
• Managed child sessions start without asking by default and consume ACUs; each CLI subagent adds its own usage on top of the parent's.
• subagent_general runs on the parent's model, so fan-out on a premium model multiplies spend.
• The Devin MCP server needs a service-user API key with a cog_ prefix; legacy apk_ keys are not supported, and Enterprise keys need an X-Org-Id header.
• Dynamic Workflows are off for enterprise accounts until an admin enables them.
• On Windows, the PowerShell install command must be run in PowerShell, not Git Bash or CMD.
• The intro page warns the docs may be out of date in places.

Yes, in both directions. Devin can use stdio, SSE and HTTP MCP servers, and Cognition hosts an official Devin MCP server at mcp.devin.ai that lets other MCP clients manage sessions, playbooks and knowledge.

A coordinator session starts managed child sessions, each in its own VM, or Devin writes a dynamic workflow script that fans work out to many sessions. The local CLI also has background subagents.

The pricing page lists Free at $0, Pro at $20 per month, Max at $200 per month, Teams at $80 per month plus $40 per month per full seat, and Enterprise on request (read 2026-09-30). Extra usage beyond quotas is bought at API pricing.

The docs describe an open-source Devin Handoff plugin that lets Claude Code, Codex, Cursor and other agents hand a task to a cloud Devin session through the Devin API.