How to Use Devin

Devin by Cognition is a managed autonomous software engineer. You give it a task in Slack, Linear, GitHub, or its own web UI, and it works in its own cloud workspace — planning, editing, running tests, opening PRs — reporting back when it’s done or stuck. It’s the most hands-off coding agent on the market, and it’s priced accordingly.

What Devin actually is

Devin is a fully hosted agent. There’s no local install, no Docker image, no model to pick. Each “session” spins up a fresh cloud VM with a browser, a shell, and an editor. Devin drives all three. It has long-running memory per workspace, learns from corrections, and can run tasks in parallel. The tradeoff: you have less control than with OpenHands or Claude Code, and you pay for Cognition’s infrastructure on top of inference.

Setting it up

Sign up at app.devin.ai, pick a plan (the Core plan starts around $20/mo for pay-as-you-go ACU credits; Team is $500/mo with a credit bundle), and connect your integrations:

# Inside Devin's settings:
# 1. Connect GitHub (repo access, PR permissions)
# 2. Connect Slack or Linear (task intake)
# 3. Add any secrets the agent needs (DB URLs, API keys) to the Secrets vault
# 4. Write a "Devin Guidelines" doc describing your repo conventions

That Guidelines doc is load-bearing. Devin reads it before every task, the same way Claude Code reads CLAUDE.md.

Your first session

Click New session and write a brief: what you want, which repo, any constraints. Devin produces a plan — read it before approving. A good first task is something like “upgrade the project from React 18 to React 19, update any deprecated APIs, keep the test suite green.” Watch the VM stream while it works. You can intervene at any time with a chat message.

A realistic workflow

Devin shines for async, multi-step work. Good fits: dependency bumps, adding telemetry across a codebase, writing migration scripts, filling out test coverage, triaging Sentry errors into PRs. Kick off two or three sessions in parallel from Slack, get a coffee, come back to draft PRs ready for review. Treat each PR like a junior’s work — read the diff, run it locally, and leave review comments. Devin responds to PR comments and pushes fixes the same way a human would.

Gotchas and limits

ACU (agent compute unit) burn is the main surprise. A meandering task can eat $10–20 of credits before you notice. Set budget limits in settings and cap session time. Devin is slower per task than running Claude Code locally — you’re paying for the hands-off part, not speed. It also fails quietly on tasks that require ambiguous product decisions; if the “right” answer needs a human call, it will guess and you’ll be rewriting the PR anyway.

When NOT to use it

Don’t use Devin for live, interactive coding — the latency makes it miserable. Don’t use it when you need to keep source on-prem without exception; while Cognition has enterprise options, default sessions run in their cloud. And don’t use it for a one-line bug fix — you’ll spend more time writing the brief than fixing it yourself. For cheaper self-hosted agents see our OpenHands guide; for in-editor flow, Cursor.