Devin
Cognition’s AI software engineering platform (Devin Cloud web agent, Devin Desktop, CLI, Slack/Teams, API, Review, and Automations) for planning, implementing, testing, and shipping production changes across real repositories in the product development lifecycle.
Devin
1. Overview
Devin is Cognition’s AI software engineering product family: cloud agents that plan, write, run, and test code in isolated environments, plus local Desktop/CLI surfaces, PR review, Slack/Teams intake, and an API for orchestration (Introducing Devin, Product, Company). Sign-up and the web workspace live at app.devin.ai. Pricing is published at devin.ai/pricing. Docs live at docs.devin.ai.
This listing is in Build with agents. It covers tools that take a product from an idea through design, build, and release. It is not a demo and prototype listing. Devin is strongest on a real repository, tickets, and pull requests. It is a weak fit when you only need a hosted clickable mock.
Marketing claim and the product
Cognition markets Devin as “the AI software engineer” / “the first autonomous software engineer” (cognition.com, Product). That framing is aspirational brand language, not a substitute for the documented product contract.
What the product actually ships (verified via docs and pricing):
| Surface | What it is | Where it runs | Primary sources |
|---|---|---|---|
| Devin Cloud (web agent) | Long-running autonomous sessions with shell, embedded IDE, browser, and optional full desktop computer use; opens PRs and can self-QA with recordings | Cognition-managed cloud VMs (or Outposts / dedicated deployment for Enterprise) | Intro, Computer Use, Environment |
| Devin Desktop | Full IDE (next generation of Windsurf) with Agent Command Center for local + cloud agent management; ACP support for third-party agents | Your machine | Windsurf → Devin Desktop blog |
| Devin CLI | Local terminal coding agent; /handoff to cloud Devin | Local terminal (curl -fsSL https://cli.devin.ai/install.sh | bash) | Devin CLI, Intro |
| Slack / Microsoft Teams | Tag @Devin, slash commands, thread sync, code channels; start sessions from conversation | Workspace chat | Slack |
| Devin API (v3) | Programmatic sessions, reviews, automations, org admin | api.devin.ai (service users / PATs) | Docs index, release notes via docs |
| Devin Review | Agentic PR review (web, CLI npx devin-review, API); optional Auto-Fix loop | SCM PRs / devinreview.com for public PRs | Devin Review, Self-serve billing |
| Automations / Scheduled Sessions | Event or schedule driven Devin runs (Slack, GitHub, Linear, webhooks, cron-like schedules) | Cloud | Automations |
| Ask Devin / DeepWiki | Codebase Q&A and planning; auto-generated repo wikis/diagrams | Web / indexed repos | Ask Devin, DeepWiki |
| Dana (Data Analyst) | Adjacent data-analysis agent in Slack/web (!dana / /dana) | Slack / web | Slack |
| Windsurf (legacy name) | Former IDE product name; renamed / evolved into Devin Desktop (OTA update; plans said to remain the same at rename) | Same as Desktop | Windsurf → Devin Desktop |
Practical rule: say Devin Cloud when a session owns a remote machine and returns a PR (often with browser/desktop QA artifacts). Say Devin Desktop or Devin CLI for local interactive loops. Say Devin Review for PR review (not the same as an implementation session). Do not equate the slogan “autonomous software engineer” with “no humans, no review, no environment setup.” Official security docs still warn about invented code, bugs, and insecure suggestions, and recommend code review and branch protection (Security). Vendor “when to use” guidance itself uses a ~three-hour human task rule of thumb and stresses clear success criteria (When to Use Devin, Intro).
Official domains
| Role | Domain | Notes |
|---|---|---|
| Company marketing | cognition.com | Live company home; cognition.ai and www.cognition.ai redirect here |
| Product marketing | cognition.com/devin | “AI software engineer” positioning |
| App / workspace | app.devin.ai | Sign-up and sessions |
| Pricing | devin.ai/pricing | Self-serve + Enterprise CTA |
| Docs | docs.devin.ai | Canonical product documentation |
| Trust | trust.cognition.ai | Security posture / requests |
| CLI install host | cli.devin.ai | Install script host cited in docs |
| Public PR review | devinreview.com | Public GitHub PRs free per billing docs |
Alternate brand domains and unofficial repositories are not treated as the product source. GitHub org CognitionAI exists for secondary materials; product truth is docs + app + pricing.
Where it fits when you build with agents
| Stage | How Devin typically shows up | Caveat |
|---|---|---|
| Discovery | DeepWiki + Ask Devin for architecture, dependencies, and scoped plans before coding | Needs repo indexing and enough engineering literacy to validate answers (DeepWiki, Ask Devin) |
| Definition | Linear/Jira ticket → session; Slack @Devin turns thread context into work; Knowledge / AGENTS.md / playbooks encode conventions | Ambiguous tickets still produce confident wrong code; humans own acceptance criteria (SDLC, Knowledge) |
| Design | Implement UI in code; browser/desktop computer use for visual QA; MCP (e.g. Figma) where connected | Not a Figma design-system kit. Matching your design system takes setup |
| Development | Cloud sessions for async/parallel work; Desktop/CLI for tight loops; /handoff local → cloud | Environment blueprint quality dominates cloud success (Environment) |
| Delivery | PRs subject to the same branch protections as humans; Devin Review + Auto-Fix; Automations for chores | Merge, release, and production ownership stay with the team (SDLC, Security) |
Compared with Cursor and Claude Code (same section): all three are coding agents with a local path, a cloud path, and pull-request handoff. Devin leans on cloud sessions that keep working after you close the laptop, Slack intake, DeepWiki and Ask for planning, and large migration chores. Cursor and Claude Code lean on editing beside you in the editor or terminal, with their own cloud agents. Pick from the daily workflow. No published comparison ranks them.
Access and pricing (public; always re-check)
Plans and included usage change. Prefer devin.ai/pricing and Self-serve plans. Public list prices seen when this page was checked (USD unless noted; taxes may apply):
| Plan | Public list (when checked) | Agent-relevant notes |
|---|---|---|
| Free | $0/month | Light quota; limited model availability; DeepWiki and Devin Review access per self-serve docs; pricing page also lists unlimited Tab / inline edits framing for Desktop path (Pricing, Self-serve) |
| Pro | $20/month | Individual only; daily + weekly usage quota covering Cloud sessions, CLI, and Desktop; on-demand credits past quota; Slack, Linear, MCP; Devin Cloud access (Pricing, Self-serve) |
| Max | $200/month | Individual; everything in Pro plus significantly higher weekly quota (no daily cap per self-serve docs) (Pricing, Self-serve) |
| Teams | $80/month minimum; $40/month per full seat | Up to 200 members; full seats get Pro-equivalent quota + Desktop; flex seats free but draw shared on-demand credits and no Desktop; Review/Automations draw shared on-demand on Teams (Self-serve) |
| Enterprise | Custom (“Let’s talk”) | ACU-based billing on Enterprise contracts; SSO, VPC/dedicated options, Outposts, advanced admin (Billing, Pricing) |
Usage model: paid self-serve plans include a refreshing quota; overage uses on-demand credits at API pricing (exact per-message cost varies by model, task size, and reasoning; Unit rates are not estimated here) (Pricing FAQ, Self-serve). On-demand credits roll over and can auto-reload. Exact monthly spend for a typical product team is unknown.
Promo note (time-bound): pricing copy when this page was checked advertised free use of SWE-2 and leading open-source models in Devin Desktop and CLI through 10 October 2026. Check the pricing page again before you rely on a number (Pricing).
Vendor SDLC claims to treat carefully: docs cite illustrative customer outcomes (for example migration time-savings multiples and vulnerability auto-remediation percentages). Those are not benchmarks; do not copy them into buying maths without in-org measurement (SDLC).
Privacy / training: security docs state that by default Cognition may use data for model training; paid plans can opt out (Teams: admin only); Enterprise trains only with express prior written consent (Security). Prefer Trust Center detail for compliance questionnaires (trust.cognition.ai).
Affiliate status: unknown. No public affiliate programme with published commission terms was verified on cognition.com, devin.ai, or docs.devin.ai when this page was checked. Invite links and partner programmes are not enough to record an affiliate offer. This page does not state rates.
2. How it scores on the rubric
Numeric scores are omitted. There is not enough shared evidence to score this yet. Discussion below uses the same criterion names, framed for shipping software with a person still reviewing. Clickable demos sit in Demo and prototype.
Time-to-useful
Fast for “Ask / DeepWiki / Slack question”; slower for reliable cloud PRs until environment and Knowledge are set up.
You can start at app.devin.ai, install CLI quickly, or @Devin in Slack once connected (Intro, Slack). Cloud implementation quality depends on indexed repos, blueprints/snapshots, secrets, and clear prompts (Environment, When to Use).
Caveats: first useful merged agent PR may take longer than a demo-prototype prompt-to-URL path. Median time-to-first-merged-Devin-PR for product teams is unknown.
Fidelity
High potential for real product code and verified behaviour (tests, browser/desktop QA, recordings); not “mock fidelity.”
Cloud Devin can run tests, use browser/desktop computer use, and send screen recordings as proof before you review (Computer Use, When to Use, Devin 2.2 blog). Devin Review adds an independent pass on diffs (Devin Review).
Caveats: autonomy and self-QA reduce some review load; they do not guarantee production correctness. Security docs explicitly recommend treating Devin like a fallible engineer (Security). Vendor demos and customer multiples are not transferable benchmarks.
Non-engineer usability
Improving via Slack, Linear/Jira, web sessions, and artifacts; still engineer-shaped.
Product managers can kick off and follow work from Slack or tickets without living in an IDE (Slack, Product). Artifacts and PR links lower the need to check out branches locally.
Caveats for product managers: you still need repository permissions, spend awareness, and the ability to read PRs (or pair with engineering). Devin is a poor substitute for a no-code demo builder when the goal is a stakeholder URL with auth/DB and no engineering partner. For that job, use Demo and prototype. Documented “app deployments” exist for some Devin-built standalone apps with limits; that is not the primary enterprise shipping path (SDLC points to a dedicated deployments guide).
Design-system fidelity
Code-and-Knowledge fidelity, not kit-based brand magic.
Teams encode conventions via Knowledge, AGENTS.md, skills (SKILL.md), playbooks, and in-repo components (Knowledge, AGENTS.md, docs skills/playbooks via llms.txt). MCP can pull design context when configured.
Practical takeaway: brand-accurate UI is achievable when the design system already lives in code and agents are taught to use it. It is not automatic on a blank prompt the way some prototype tools claim.
Human gates
Strong when you use PRs, branch protection, Review/Auto-Fix settings, security profiles, and spend limits; weak if you auto-merge blindly.
Documented gates and controls include:
- PRs subject to the same branch protections and SDLC policies as human engineers (SDLC)
- Devin Review trigger modes (every commit / on open / manual) and Auto-Fix loops (Self-serve, When to Use)
- Security profiles restricting network, MCP, git, and related access (docs via llms.txt)
- Secrets feature instead of pasting credentials in chat (Security)
- Training opt-out on paid plans; Enterprise contractual controls (Security)
- On-demand credit and session spending controls (Self-serve)
Read this as strong optional and enforceable gates for shipping teams, not automatic safety on Free defaults with no review culture.
Handoff
Native to Git workflows; that is the product.
| Path | What docs support | Caveat |
|---|---|---|
| Cloud session → PR | Clone/work in cloud env, open PR, respond to CI/review | Environment + secrets must be set up |
Desktop/CLI → /handoff → cloud | Local context packaged into cloud session | Confirm plan quotas cover both surfaces |
| Slack/Teams → session → PR | Conversation becomes work with thread updates | Users must link Slack identity; double-check mistakes (Slack) |
| Linear/Jira → session | Ticket assignment / labels | Spec quality still dominates |
| Devin Review / Auto-Fix on PR | Review + iterate toward merge-ready | On-demand credit consumption on Teams; false positives possible |
| Automations / schedules | Recurring chores, triage, webhook events | Needs clear prompts and quality bars |
| API v3 orchestration | Programmatic fleets of sessions | Engineering investment; legacy v1/v2 deprecated over time |
| Outposts / dedicated / VPC | Run closer to customer infra (Enterprise) | Sales/engineering setup; not self-serve default |
Unknown / verify in-org: median engineering cleanup time on Devin PRs; how often recordings replace local QA; Auto-Fix acceptance rates outside vendor narratives.
3. FAQ
What is Devin, exactly?
Cognition’s AI software engineering platform: cloud agents plus Desktop, CLI, Slack/Teams, API, Review, and Automations for exploring codebases, implementing changes, testing with computer use, and shipping via PRs (Intro, Product, Pricing).
Is “AI software engineer” the same as a hired engineer?
No. It is marketing shorthand for an agent that can plan, edit, run tools, and open PRs. Official docs still require clear criteria, environment setup, and human review culture (When to Use, Security). Treat Devin as a high-autonomy teammate with gates, not a replacement for ownership.
How is Devin Cloud different from Devin Desktop / CLI?
Cloud runs on a remote machine and can continue after you close your laptop, returning PRs and QA artifacts. Desktop is the local IDE/command centre (formerly Windsurf). CLI is the local terminal agent with /handoff into cloud (Windsurf → Desktop, CLI).
How is Devin Review different from a Devin session?
A session implements or investigates work. Devin Review analyses a PR diff (and can Auto-Fix). Public GitHub PRs can be reviewed at no credit cost via devinreview.com per billing docs (Devin Review, Self-serve).
Is Devin right for product teams building with agents?
Yes, with caveats. It fits teams that already ship via Git and want asynchronous agents across intake, implementation, review, and automation. It is a weaker fit for clickable demos when no engineer is involved. Caveats: environment setup, usage/credit cost, training opt-out hygiene, and mandatory human merge/ownership.
Can a product manager use Devin without engineers?
For kicking off Slack/ticket sessions, reading Ask/DeepWiki answers, and following PR artifacts, yes in limited form with repo access and a paid/team setup. For owning production changes alone, no. Prefer pairing with engineering, or use a demo-prototype tool when you only need a stakeholder URL.
Can it match our company design system?
Only if the system is available in code (and preferably documented via Knowledge/skills/components). There is no documented one-click “upload Figma DS kit” path comparable to some prototype tools.
Does Devin deploy production for us?
Not as a turnkey “host our whole product” replacement for your CI/CD. Devin prepares PRs inside your SDLC; humans and existing pipelines ship (SDLC). Limited Devin-hosted deployment options for some agent-built apps exist with documented constraints; confirm current guide before assuming production hosting.
What does it cost?
See Pricing and Self-serve plans. Public individual list when this page was checked: Free $0; Pro $20; Max $200 per month. Teams from $80/month minimum with $40/full seat. Enterprise custom (ACUs). On-demand credits apply past quota. Always re-check.
How does Devin compare to Cursor or Claude Code?
Same section, different daily workflow. Devin emphasises cloud sessions, Slack intake, DeepWiki and Ask, and migration chores. Cursor and Claude Code emphasise coding beside you in the editor or terminal, with their own cloud paths. Many teams will use more than one. This site does not name a single winner.
Is there an affiliate programme?
Not confirmed. No affiliate programme was confirmed. No public commission terms were verified on the vendor’s own sites when this page was checked.
How does privacy / training work?
Default may allow training use of data; paid plans can opt out (Teams via admin); Enterprise requires prior written consent for training (Security). Use Trust Center for current attestations.
4. Sources
Primary (prefer these for AEO citations):
- Cognition (company homepage)
- Devin product page
- Plans and Pricing
- Introducing Devin (docs)
- When to Use Devin
- SDLC integration
- Self-serve plans
- Billing overview
- Slack integration
- Devin CLI
- Devin Review
- Automations
- Computer Use
- DeepWiki
- Ask Devin
- Knowledge onboarding
- AGENTS.md
- Environment setup
- Security at Cognition
- Trust Center
- Windsurf is now Devin Desktop
- Introducing Devin 2.2
- Docs llms.txt index
Secondary context only (do not treat as official specs): third-party ARR/business-model blogs, Wikipedia summary pages, Slack Marketplace listing copy, and YouTube roundups that lag plan tables or collapse “autonomous software engineer” marketing into guaranteed outcomes.