Migrating from Devin Desktop (formerly Windsurf)

Migrating from Devin Desktop, the editor Cognition renamed from Windsurf in June 2026, means translating each .windsurf/rules/ file by its trigger into instruction files, path-scoped rules or skills, turning .windsurf/workflows/ into skills, and re-adding MCP servers per tool. The migration is done when a behavioural check proves Claude Code, Codex or Cursor applies the same rules.

Your team adopted Windsurf last year. The editor has since updated itself into Devin Desktop, the repository holds 14 files in .windsurf/rules/ with four different trigger values and six slash-command workflows in .windsurf/workflows/, and half the team now wants Claude Code or Codex. Concatenating the rule files into one AGENTS.md looks like a five-minute job, until the agent starts applying your Go-only rules to TypeScript and ignores the ones that were only ever loaded on demand.

This page is for the developer who does the move and the tech lead who has to keep the team’s rules intact while people switch at different speeds. It covers the repository side of the move; the product comparison lives in Cursor, Claude Code, and Codex vs Devin Desktop (formerly Windsurf).

What we could verify about the rename

The rename to Devin Desktop is reported by Creeta (June 2026, secondary source); Cognition’s own pages were unreachable from our environment on 2026-09-26, so this page makes no claim about which Windsurf features Devin Desktop still ships or how it now stores configuration. The .windsurf/rules/ trigger values below come from public repositories (GitHub code search, 2026-09-26), and some repositories already keep rules in .devin/rules/. The inventory step finds both.

What this migration guide covers

• A decision table for whether to switch editor at all, or add a terminal agent beside Devin Desktop
• A trigger-by-trigger map from Windsurf rules to AGENTS.md, CLAUDE.md, path-scoped rules and skills
• Four copy-paste prompts: inventory and translate rules, convert workflows to skills, prove the rules apply, and run a first delegated task with evidence
• MCP commands for Claude Code and Codex, and the verification gates that make the move provable
• The failure modes of a half-migrated team and how to recover from each

Do you need to leave the editor to switch agents?

Not necessarily. Decide what you are replacing before you move any file.

Situation	What to do
You like the editor and want a stronger terminal agent	Keep Devin Desktop and run Claude Code or Codex beside it in a terminal (or in the editor’s terminal, if your build still has one). OpenAI’s Codex README lists Windsurf among the editors its IDE extension supports (checked 2026-09-26). Translate the rules so both agents read them.
You want an IDE-first agent with inline completion	Move to Cursor, the only one of the three that is an editor. Its models and prices could not be re-verified on 2026-09-26, so check cursor.com before you buy.
You want agents in CI, scheduled runs or headless scripts	Add Claude Code (claude -p) or Codex (codex exec). Both run without an editor; if you stay with Cognition, compare them with Devin CLI in the coding-agent landscape.
You depend on a model only Cognition offers, such as SWE-2 (in Devin Desktop and Devin CLI; launched 2026-09-10 per aipricing.guru, secondary)	Keep that work in Devin Desktop until your own evals show a replacement matches it. The models hub lists what each tool runs.
You are comparing more than the big three	Start with the coding-agent landscape, which covers Devin and Devin CLI as well.

Start on each tool’s default model: Claude Opus 5.5 in Claude Code (from v2.1.280 on the latest release channel) and GPT-6 Astra in Codex. Tune effort before you switch model, and switch only when your evals say so.

How do Windsurf rule triggers map to Claude Code, Codex and Cursor?

A Windsurf rule file carries a trigger in its frontmatter, and the trigger decides when the agent sees it. The target depends on the trigger, not on the file. Flattening every file into one instruction file is the most common migration mistake, because it turns on-demand rules into always-on noise.

Windsurf trigger	Claude Code	Codex	Cursor
always_on	CLAUDE.md (first line @AGENTS.md)	Root AGENTS.md	A rule set to apply always
glob with globs:	.claude/rules/NAME.md with a paths: list in its frontmatter	AGENTS.md in the directory the glob covers	A rule scoped to the same globs
model_decision with description:	Skill in .claude/skills/NAME/SKILL.md; its description decides when it loads	Skill in .agents/skills/NAME/SKILL.md	Skill in .agents/skills/NAME/ (reuse the Codex copy), or a rule the agent requests by description
manual	Skill with disable-model-invocation: true, invoked as /NAME	Skill you invoke with $NAME or pick from /skills	Skill in .agents/skills/NAME/, or a rule you mention explicitly
.windsurf/workflows/NAME.md (a /NAME workflow)	Skill	Skill	Skill
Auto-generated Memories	Review, then promote durable ones to AGENTS.md; Claude Code’s auto memory takes over from there	Promote to AGENTS.md	Promote to a rule

A glob that maps to one directory becomes a nested AGENTS.md there for Codex. An extension or cross-directory glob such as **/*.go or **/*.test.ts has no Codex equivalent, because Codex scopes AGENTS.md by directory, not by file pattern (codex-cli 0.157.1, checked 2026-09-26): write it as a conditional line in the root AGENTS.md (“When editing *.go files: …”) or as a skill, and test it with question 3 of the proof prompt below.

The Claude Code and Codex locations were checked against the installed CLIs (Claude Code 2.1.283, codex-cli 0.157.1) on 2026-09-26. Claude Code reads AGENTS.md on its own only when there is no CLAUDE.md, and only from v2.1.277 (first-party sessions) or v2.1.281 (Bedrock, Google Cloud, Foundry and LLM gateways), both on the latest channel only, which is why the table routes it through an @AGENTS.md import that works on both channels. Cursor’s rule file format could not be re-verified on 2026-09-26: create one rule through Cursor’s own rules settings, then have the agent copy that format. The same goes for Cursor’s skill locations: a Cursor docs snippet lists .agents/skills/ and also .claude/skills/, but we could not re-verify it on 2026-09-26, so confirm that Cursor lists the migrated skills before you rely on them.

Migrate the repository in five steps

1. Inventory everything Windsurf-shaped. Run this in a terminal at the repository root. The first command lists rule, workflow and instruction files in every directory, not only the root:

   find . -path ./node_modules -prune -o -type f \( -path '*/.windsurf/*' -o -path '*/.devin/*' -o -name '.windsurfrules' -o -name 'AGENTS.md' -o -name 'CLAUDE.md' \) -print
   find . -path ./node_modules -prune -o -type f \( -path '*/.windsurf/rules/*' -o -path '*/.devin/rules/*' \) -exec grep -h '^trigger:' {} + | sort | uniq -c

   The second command counts rules per trigger across every directory, including nested .windsurf/ folders in a monorepo, which tells you how many land in each column of the table above. Global rules, Memories and MCP servers live in the editor’s settings on each engineer’s machine, not in git. Export them from each machine before anyone uninstalls the editor.

2. Translate rules by trigger. Run the first prompt below in the agent you are moving to. It writes the target files and a mapping table you check instead of reading every rule.

3. Turn workflows into skills. Run the second prompt. A Windsurf workflow is a named, repeatable procedure, which is exactly what an Agent Skill is. If the team runs both Claude Code and Codex, keep one copy of each skill and confirm discovery in each tool: type / in Claude Code, and run /skills in Codex.

4. Re-add MCP servers per tool. MCP configuration does not transfer between editors. Add each server again and read secrets from environment variables, never from the command line. The example below uses the Playwright MCP server (@playwright/mcp).

   Claude Code

   Run in a terminal. --scope project writes .mcp.json at the repository root, so the whole team gets the server from git.

   claude mcp add --scope project playwright -- npx @playwright/mcp@latest

   Codex

   Run in a terminal. Codex writes the entry to ~/.codex/config.toml, so each engineer runs it once.

   codex mcp add playwright -- npx @playwright/mcp@latest

   Cursor

   Add the server through Cursor’s MCP settings with the command npx @playwright/mcp@latest. Check the configuration format on cursor.com/docs/mcp; it was not re-verified on 2026-09-26.

   Which servers are worth adding at all is covered in essential MCP servers.

5. Prove the rules apply, then commit. Run the third prompt in a fresh session. Commit the new files only when every rule it tests is answered from the right source. Keep .windsurf/rules/ in place until the last engineer has switched.

Copy-paste prompt: translate Windsurf rules by trigger

Migrate this repository's Windsurf / Devin Desktop rules to the agent you are running in.

1. Read every file under .windsurf/rules/ and .devin/rules/, plus .windsurfrules if it exists. For each file, record its trigger (always_on, glob, model_decision, manual) and any globs or description.
2. Map each file by trigger:
   - always_on -> root AGENTS.md (create it if missing). If you are Claude Code, make CLAUDE.md start with the line @AGENTS.md.
   - glob -> a path-scoped rule (Claude Code: .claude/rules/<name>.md with a paths: list; Codex: an AGENTS.md inside the directory the glob covers, or, for an extension or cross-directory glob, a conditional line in the root AGENTS.md; Cursor: a rule scoped to the same globs).
   - model_decision -> a skill whose description is the rule's description.
   - manual -> a skill invoked by name only (Claude Code: add disable-model-invocation: true to its frontmatter; Codex: invoke it with $NAME).
3. Rewrite Windsurf-specific wording (Cascade, Flows, Supercomplete, Memories) into tool-neutral instructions. Do not change any convention.
4. Run every build, test, lint and type-check command the rules mention. Keep the ones that succeed; list the ones that fail.
5. Do not delete .windsurf/ or .devin/. Do not change source code.

Finish with a table: original file, trigger, new location, and anything you dropped with the reason.

Copy-paste prompt: convert Windsurf workflows into skills

Convert each file in .windsurf/workflows/ into an Agent Skill.

For each workflow:
1. Create a skill directory named after the workflow in kebab-case (Claude Code: .claude/skills/<name>/; Codex: .agents/skills/<name>/; Cursor: .agents/skills/<name>/, the same copy Codex uses), containing SKILL.md.
2. Write frontmatter with name (at most 64 characters) and a description of at most two sentences that says when to use the skill, so the agent can pick it without being asked.
3. Move the workflow's steps into the body unchanged in meaning. Replace any Windsurf-only step (a Cascade action, a Windsurf UI click) with the equivalent shell command or instruction, and mark each replacement with a comment line starting "Migrated:".
4. If a step runs a script, keep the script in the skill directory and reference it by relative path.

Do not invent steps. Finish with a list of each workflow, its new skill path, and every "Migrated:" replacement for me to review.

How do you prove the migrated rules actually apply?

A mapping table shows where each rule went. It does not show that the agent reads it. Ask questions whose only correct answers come from specific rules, in a fresh session with no other context, and require the agent to cite the file.

Copy-paste prompt: prove the migrated rules apply

Answer each question using only the instruction files, rules and skills loaded in this session. For each answer, name the file it came from. If you do not know from a loaded file, answer RULE_NOT_LOADED instead of guessing.

1. Which command runs the full test suite, and which runs a single test file?
2. Which conventions apply when you edit a file under src/api/? Which apply under web/?
3. Which of those conventions do NOT apply when you edit a Markdown file in docs/?
4. List the skills available to you and, for each, one sentence on when you would use it.
5. What must be true before you call a change done?

Replace src/api/ and web/ with two directories your glob rules cover. Question 3 is the one that catches flattening: if the agent applies a Go-only or API-only rule to docs/, a path-scoped rule became global. Any RULE_NOT_LOADED answer for an always_on rule means the instruction file is not being read; see the failure modes below.

Then give the agent a real task with acceptance criteria, the kind you used to hand to Cascade, and judge it by evidence rather than by reading the diff.

Copy-paste prompt: a first delegated task that proves itself

Add rate limiting to POST /api/login: at most 5 failed attempts per email per 15 minutes, then HTTP 429 with a Retry-After header.

Acceptance criteria:
- A test sends 5 failed logins and gets 401 each time, then a 6th and gets 429 with Retry-After.
- A successful login resets the counter (test it).
- Existing auth tests still pass unchanged; do not edit them.
- Type check, lint and the full test suite pass.

Plan first and wait for my approval. Work on a new branch named after your tool (claude/login-rate-limit, codex/login-rate-limit or cursor/login-rate-limit). Open a pull request whose body maps each criterion to the test or command that proves it, with the output. If a criterion cannot be met, stop and explain.

Claude Code

Start in plan mode so nothing is edited before you approve the plan. Shift+Tab cycles modes in a running session, and /rewind (Esc Esc) restores a checkpoint if the agent goes wrong.

claude --permission-mode plan

Codex

Run the prompt interactively so you can approve the plan before any edit.

codex

codex exec cannot run this prompt as written: it is non-interactive, so it cannot wait for approval, and a sandboxed run may have no network to open a pull request. For a headless run, delete the “Plan first and wait for my approval” sentence and replace “Open a pull request” with “Write the evidence to EVIDENCE.md and stop”. Run Codex non-interactively covers the rest of a headless setup.

Cursor

Paste the prompt into the agent with Plan Mode on; Cursor’s docs described Plan Mode on 2026-08-28 as creating “detailed implementation plans before writing any code”. Approve the plan, then let the agent run the tests.

How do you verify migrated agent work without reading every line?

The editor changes; the gates do not. Put these in place before the first migrated pull request merges.

• Acceptance criteria before code. Every delegated task names its criteria and the command that proves each one, as in the prompt above. See acceptance criteria an agent can check.
• Required CI checks. Type check, lint and tests block the merge whichever agent or editor wrote the code.
• An evidence bundle in each pull request. Each criterion maps to a test and its output, so the reviewer checks the mapping. See the evidence bundle.
• A rules check in CI. Save the proof prompt as .github/rules-check.md, rerun it headlessly after any change to the instruction files, and fail the job when an answer is the sentinel token: claude -p "$(cat .github/rules-check.md)" > rules-check.txt && ! grep -q 'RULE_NOT_LOADED' rules-check.txt. A unique token avoids false failures from a correct answer such as “those rules do not apply to docs/”. The headless run needs ANTHROPIC_API_KEY as a CI secret. On a Codex-only team, the equivalent step is codex exec "$(cat .github/rules-check.md)" > rules-check.txt && ! grep -q 'RULE_NOT_LOADED' rules-check.txt, with CODEX_API_KEY as the secret. The project must be trusted in the runner’s Codex configuration, because since codex-cli 0.150.0 an untrusted project does not supply AGENTS.md and every answer comes back RULE_NOT_LOADED. Run Codex non-interactively covers authentication and the rest of the CI setup.
• Protected tests. Agents must not weaken a test to pass it; a rule in AGENTS.md plus CODEOWNERS on the test directories keeps changes visible. See protecting the test oracle.
• Human sign-off where tests cannot reach. The engineer who delegated the task approves the evidence and reads the diff only for auth, migrations, billing and public API changes. Reviewing agent pull requests has the checklist.

What breaks when a team migrates off Devin Desktop?

Every rule became always-on. Someone concatenated .windsurf/rules/*.md into AGENTS.md, and the agent now applies framework rules to the wrong files while the context fills with rules it does not need. Recovery: rerun the translation prompt from the original files, and check question 3 of the proof prompt.

Claude Code does not see the rules. Engineers on the stable channel (v2.1.274 on 2026-09-26) do not get the AGENTS.md fallback, and a CLAUDE.md without the import hides AGENTS.md on every channel. Recovery: start every CLAUDE.md with @AGENTS.md, and add a CI step that fails otherwise (head -n 1 CLAUDE.md | grep -qx '@AGENTS.md').

Codex ignores AGENTS.md in a fresh clone. Since codex-cli 0.150.0, untrusted projects do not supply project-level AGENTS.md instructions. Recovery: trust the project when Codex asks, then rerun the proof prompt.

Rules drift between .windsurf/ and the new files. During the transition someone updates the test command in one place only. Recovery: pick the new files as the source of truth on day one, add a note at the top of each .windsurf/rules/ file pointing to its new location, and delete .windsurf/ when the last engineer has switched.

MCP servers are missing. The new agent cannot reach the database or browser the old setup had, because MCP configuration lived in the editor’s settings. Recovery: rebuild the list from each engineer’s export in step 1, and commit Claude Code’s .mcp.json so the next person gets it from git.

Memories vanish. Knowledge the editor had generated on its own never reached git. Recovery: before uninstalling, have each engineer review their Memories and promote the durable ones (commands, conventions, known traps) into AGENTS.md through a normal pull request.

A workflow’s shortcut does nothing. Typing /deploy-preview in Claude Code (or $deploy-preview in Codex) does nothing because the skill’s description or directory is wrong. Recovery: check discovery with / in Claude Code or /skills in Codex, and fix the name and description frontmatter.

Where to go next after Devin Desktop

Cursor, Claude Code, and Codex vs Devin Desktop (formerly Windsurf) The product comparison behind the decision table.

Migrating between Codex, Cursor, and Claude Code Keep one set of instructions across the three tools.

Project conversion playbook Make the repository agent-ready beyond the rules.

Team onboarding and adoption, loop by loop Extend the migration from one engineer to the whole team.

To install the new agent, start with the Claude Code quick start, the Codex quick start or the Cursor quick start. For a team running two agents during the switch, adding Claude Code or Codex alongside GitHub Copilot shows the same pilot and metrics pattern.