GILD verdict
Choose Devin Desktop if your team will switch editors to run Cognition's agents, third-party agents, and Devin cloud sessions from one place.
Budget for Enterprise if you need SSO, and set agent permission rules before usage grows.
Devin Desktop (formerly Windsurf)
AI Coding Agent by Cognition
+17 MixedFit and limits
Who it fits
- Engineers will work in a VS Code-derived editor. VS Code and Cursor settings import during onboarding or from the command palette.
- You want to run Devin Local next to third-party agents over ACP, such as Codex CLI, Claude Agent, or Gemini CLI, and track every session on one board.
- You plan to delegate longer tasks to Devin cloud sessions, which self-serve plans fund from the same quota as local agent work.
- Admins want to enforce agent behavior centrally: allow, ask, and deny rules, MCP and model allowlists, and sandbox mode on Enterprise.
- Token spend per task matters to you. On a 14-file refactor, Maksim Danilchenko measured about 18,000 tokens for Devin Local against 24,000 for Cascade, in line with Cognition's claim of up to 30% better efficiency.
Where it falls short
- SSO and role-based access are Enterprise features. New Teams plans no longer include SSO.
- Cascade was removed on Sep 8, 2026: Devin Local is the only built-in agent, and Quick Review works with it.
- The Windsurf JetBrains plugin is in maintenance mode, and Cognition points JetBrains users to Devin over ACP instead.
- Teams flex seats do not include Devin Desktop, so every editor user needs a US$40 full seat.
- Devin Local is a new agent with new habits. Danilchenko warns reviewers not to assume it ran the tests after a refactor.
Implementation reality
Choose the agent and its permissions
New tabs default to Devin Local, which replaces Cascade's auto-execution levels with allow, ask, and deny rules and asks before every MCP tool call by default. On Enterprise plans an admin must turn Devin Local on, and can force sandbox mode with organization-wide domain rules.
Put a ceiling on spend
Self-serve plans include a daily and weekly quota drawn down per token at each model's rate, and extra usage is billed at API list prices. On Teams, on-demand credits are shared across members, and admins set auto-reload thresholds and default session spending limits. Newer Enterprise contracts are billed in ACUs.
Decide on data-retention models
Some Anthropic and OpenAI models require the provider to retain data for up to 30 days for trust and safety checks, and Teams and Enterprise admins choose whether to allow them. Web search is off by default for Enterprise teams.
Pick the Quick Review model
Quick Review offers SWE-check, Cognition's own review model, free on every plan. Cognition reports that SWE-check matches Claude Opus 4.6 on its in-distribution bug evals and trails it on out-of-distribution ones. GPT 5.5 and Opus 4.7 reviews draw quota per token, and Enterprise admins can limit which review models engineers see.
Watch cloud sessions in the quota
Devin cloud sessions started from the editor draw on the same shared quota and extra usage balance as local agent work. Every user can see a session's usage in Session Insights.
What to measure in your pilot
- GILD found one hands-on token count and no team-level data on Devin Local; compare reviewed changes merged per day of quota in the first month against your last Windsurf month.
- Cognition's usage pages show per-session usage but no split between local and cloud work; tag cloud sessions for a month and cap them if they take most of a seat's quota.
- GILD found no independent score for Quick Review; log which SWE-check findings reviewers act on for a month, and move risky repositories to a frontier review model if most are dismissed.
Sources
- Official documentation
- 20
- Public source
- 1
- Devin Desktop
- Devin Desktop docs: FAQ on the move from Windsurf
- Devin docs: Self-serve plans
- Devin Desktop docs: Quota-based usage
- Devin Desktop docs: Plans and usage
- Devin Desktop docs: Install
- Devin Desktop docs: Welcome to Devin Desktop
- Devin Desktop docs: Devin Local agent
- Devin Desktop docs: Cascade overview
- Devin Desktop docs: Terminal
- Devin Desktop docs: Agent Client Protocol
- Devin Desktop docs: Agent Command Center
- Devin Desktop docs: Devin in Devin Desktop
- Devin Desktop docs: Quick Review
- Devin Desktop docs: Context awareness
- Devin Desktop docs: Admin controls reference
- Devin Desktop docs: SSO and SCIM
- Devin Desktop docs: Role-based access
- Maksim Danilchenko: Devin Desktop (ex-Windsurf): Features, Pricing, What Changed (Jun 6, 2026)
- Cognition: Introducing SWE-Check: 10x Faster Bug Detection (Apr 14, 2026)
- Devin docs: Usage