← Back to catalog

tool

Devin — AI Software Engineer Cheatsheet

Set up and operate Devin, Cognition's autonomous AI software engineer: connect repositories, configure the environment snapshot, run Ask-to-Agent sessions, move work between the Devin CLI and cloud VMs, and automate recurring tasks.

  • 2026-10-03 (2 weeks ago)
  • 6 steps
  • ~100 min total
  • Devin current
  • devin
  • cognition-ai
  • coding-agent
  • cloud-agent
  • devin-cli
  • automations

Devin — AI Software Engineer Cheatsheet

One-line: Devin is Cognition’s autonomous AI software engineer — give it a scoped task and it plans, codes, tests, and opens pull requests on its own cloud VM, with a Devin CLI for local work that can hand off to the cloud.

Prerequisites

  • A Devin account — sign up at app.devin.ai (Free, Pro, Max, and Teams self-serve plans; Enterprise via Cognition)
  • Admin rights on your GitHub organization to connect repositories (or the equivalent on GitLab 15.0+, Bitbucket, or Azure DevOps)
  • Optional: a supported terminal for Devin CLI — see the terminal compatibility docs in Reference

Installation

Devin has two surfaces. Cloud Devin needs no install — everything starts at app.devin.ai. For the terminal:

# Install Devin CLI — a local agent with deep Devin Cloud integration
curl -fsSL https://cli.devin.ai/install.sh | bash

# Restart your terminal, then authenticate
devin auth login
devin --version   # prints the installed version
devin             # start an interactive session in the current directory

Devin CLI and cloud Devin are separate tools: the CLI edits your local files, while cloud sessions get a dedicated VM with a shell, IDE, and browser. Knowledge, Playbooks, and Secrets are cloud-side features the CLI does not yet support.

Step-by-Step Setup & Optimization

Step 1 — Connect repositories and index your code

Goal: Devin can clone your repos and answer code questions via Ask Devin and DeepWikiTime: ~10 minLevel: beginner

  1. In the web app, open Settings → Connections → GitHub, click Add Connection, and grant access to All repositories or Select repositories (GitLab, Bitbucket, and Azure DevOps follow the same flow).
  2. Enable code search: go to Settings → DeepWiki → Repositories, click Add, and pick the repositories to index.
  3. Wait for each repo’s status to move from Indexing… to Indexed — a few minutes depending on size.
the repository shows Indexed under Settings → DeepWiki, and Ask Devin answers a question about it with file citations.

Index the branches your team actively develops on so Ask Devin and DeepWiki stay current.

Step 2 — Configure the environment snapshot

Goal: Every session boots a VM with your repos, toolchain, dependencies, and secrets already in placeTime: ~30 minLevel: intermediate

The docs call this the single highest-leverage thing you can do to improve Devin’s effectiveness.

  1. Start a session and ask: “Set up your environment for this repo.”
  2. Review the blueprint Devin proposes as suggestion cards — approve the ones that match your stack.
  3. Let the build finish: it produces the snapshot every session in your organization boots from.
  4. Add credentials at app.devin.ai/secrets (organization or personal scope), or per-repo in the blueprint editor’s Secrets tab — sessions reference them as environment variables like $MY_SECRET.
a fresh session runs your project’s test or lint command with zero manual setup.

Never put secret values in blueprint YAML, setup commands, or committed .env files — store them in the Secrets tab and reference them by name.

Step 3 — Run your first task: Ask → Agent

Goal: A draft pull request produced by a well-scoped sessionTime: ~15 minLevel: beginner

  1. Toggle Ask mode, select the repo(s), and scope the task — Ask Devin cites code and builds a context-rich prompt without touching code.
  2. Click Send to Devin to move into Agent mode, which executes the plan on a cloud VM.
  3. Structure prompts as directive + context + verification: what to change, which patterns or files to follow, and the exact check that proves it works.
  4. Watch progress on the session page; take over in the embedded IDE, shell, or browser if needed.
Devin opens a pull request in the repository and reports CI status on the session page.

Rule of thumb from the docs: if a task would take you three hours or less, Devin can most likely do it. Split bigger work into parallel sessions.

Step 4 — Bridge local and cloud with Devin CLI

Goal: The same task moves freely between your terminal and a cloud VMTime: ~10 minLevel: intermediate

  1. Start local work from a project directory:

    devin                            # interactive session
devin -p "summarize this repo"   # single-shot: prints and exits

  2. Start or resume cloud sessions — they keep running when you close the laptop:

    devin --cloud                      # new cloud session
devin --cloud -r                   # pick a recent cloud session to resume
devin ssh <session-id-or-url>      # shell on the session's VM

  3. Inside a session, /handoff ships the local session to the cloud (conversation context and uncommitted diff included); in a cloud session, /handoff or /pickup checks out the PR branch and continues locally. /open opens the session in the web app.

devin --version prints a version number, and /handoff returns a cloud session URL that opens in the web app.

Step 5 — Teach Devin your conventions

Goal: Durable context — skills, rules, AGENTS.md, playbooks — instead of repeated prompt remindersTime: ~15 minLevel: intermediate

  1. Add an AGENTS.md at each repo root — Devin auto-injects it into cloud sessions (automatic injection is capped at 16 KiB; see the agents-md doc).
  2. Open app.devin.ai/customize and install marketplace plugins — bundles of skills, rules, hooks, and MCP servers — at personal, organization, or enterprise scope. From the CLI: devin plugins install <source>.
  3. Save repeated task prompts as Playbooks, then attach them with a !macro or a .devin.md file at session start.
  4. Connect external tools under the MCPs tab (Datadog, Sentry, Figma, databases) so sessions can reach them.
the skill or rule appears under Customize → Skills/Rules, and a new session can invoke installed skills as /<plugin>:<skill>.

Knowledge is deprecated and being migrated to Skills in Plugins — author new context as skills and rules, not Knowledge items.

Step 6 — Close the loop: reviews, comments, automations

Goal: Devin reacts to events and finishes PRs without you in the loopTime: ~20 minLevel: advanced

  1. On any open PR in a connected repo, comment /devin fix the failing lint check — Devin starts a session and replies with a link (/devin review triggers Devin Review).
  2. Enable Devin Review with Auto-Fix so Devin responds to review comments and iterates on CI failures automatically.
  3. Build an automation (Automations → Create automation): pick a trigger — GitHub check-run failure, Slack message or reaction, Linear or Jira event, Schedule, Webhook — and an action (start a session, message a session, or run a triage monitor).
  4. Run independent tasks in parallel sessions — there is no concurrent session limit.
a test trigger — a /devin PR comment or a scheduled run — spawns a session that links back to the PR or thread.

GitHub automations fire on private repos by default. Opting a connection into public repos raises prompt-injection risk — keep trigger conditions narrow.

Best Practices

Do

  • ✅ Scope with Ask Devin first, then let Agent mode execute the generated plan.
  • ✅ Put explicit success criteria and verification commands in every prompt.
  • ✅ Keep sessions focused (XS–M in Session Insights); parallelize big projects across sessions.
  • ✅ Store credentials as Secrets on a dedicated [email protected] service account.
  • ✅ Encode team conventions in AGENTS.md, skills, and rules; standardize repeated tasks as Playbooks.
  • ✅ Check Session Insights after runs and convert mistakes into rules or skills.

Don’t

  • ❌ Paste secrets into prompts, blueprint YAML, or committed .env files.
  • ❌ Leave design decisions open-ended — make judgment calls for Devin up front.
  • ❌ Cram unrelated tasks into one session — it burns usage and degrades quality.
  • ❌ Author new Knowledge items — the feature is deprecated in favor of Skills and Rules.
  • ❌ Assume the CLI sees cloud-side features — Knowledge, Playbooks, and Secrets are web-only for now.

Quick Command Reference

CommandUse
devinStart an interactive session in the current directory.
devin -- <prompt>Start the session with an initial prompt.
devin -p "<prompt>"Single-shot: print the response and exit.
devin -c / devin -r [id]Continue the last session / resume a session.
devin auth login / status / logoutManage CLI authentication.
devin --cloudStart a Devin Cloud session from the terminal.
devin --cloud -r <url-or-id>Resume a cloud session in the terminal.
devin ssh <session>Open a shell on a cloud session’s VM.
/handoffMove the task between local and cloud — works both ways.
/cloud · /repo · /platform · /modelConfigure a cloud session before the first message (SWE-2 recommended).
/mode <name>Permission mode: normal, accept-edits, smart, plan, bypass (autonomous under --sandbox).
/plan · /askPlan-first and question-only agent modes.
/loop <prompt>Run a prompt, then auto-review the diff in a loop.
devin plugins install <source>Install a plugin into personal scope (--local for this machine only).
npx devin-review <pr-url>Run Devin Review on a PR from a local clone.
/devin <prompt>GitHub PR comment that starts a session (/devin review runs Devin Review).

Expected Outcomes

  • Repositories are connected and indexed — Ask Devin and DeepWiki answer with file citations.
  • Every session boots from a snapshot with dependencies, tools, and secrets ready.
  • The Ask → Agent workflow turns scoped prompts into draft PRs for roughly three-hour tasks.
  • Devin CLI moves work between local and cloud with devin --cloud, /handoff, and devin ssh.
  • Team conventions live in AGENTS.md, skills, rules, and playbooks — not repeated prompt reminders.
  • Automations and Devin Review Auto-Fix keep CI and review loops moving without manual steering.
  • Usage stays controlled: short scoped sessions, automatic sleep after 30 idle minutes, and Session Insights for post-task review.

Reference

Sources & deeper reading

Official docs (facts and commands):

Community sources:

Omitted areas (see docs): Devin API reference, Devin Desktop, Outposts self-hosted workers, Enterprise SSO/SCIM, macOS/Windows/Android environments, DeepWiki configuration, Security Swarm and Code Scans.