How to Reduce Devin Desktop Token Usage

You reduce Devin Desktop token usage by controlling what its agent reads on every step: the files the index retrieves, the rules loaded into every request, the MCP tools advertised each turn, and the history a long session drags along. The rename from Windsurf changed some of those defaults, so old advice is only half right.

Part of the reduce AI coding agent token usage pillar. Still on a legacy Windsurf setup or the Cascade agent? The Cascade-era levers (indexed retrieval, terminal output, long Flows) are covered in how to reduce Windsurf token usage. This page covers what's new under the Devin name.

What is Devin Desktop, and what changed for tokens?

Devin Desktop is Windsurf under a new name, but its default agent is new: Devin Local replaces Cascade for new conversations.

Cognition's Devin Desktop FAQ says the rename shipped on June 2, 2026, as a regular update: same IDE, same editor, Windsurf settings ported automatically. Plans and pricing were unchanged at the switch.

The agent is the real change. Per the Devin Local docs, new tabs start with Devin Local, the harness shared with Devin CLI. "New conversations never start on Cascade", though old Cascade conversations stay in your history.

Should you switch from Cascade to Devin Local?

Yes, for new work: Cognition built it with token efficiency as an explicit goal.

Cognition advertises that Devin Local is "significantly more token-efficient, with a greater focus on prompt caching", and that most tasks take "up to 30% fewer tokens than Cascade". That's their own comparison against their previous agent, on their tasks. It isn't an independent measurement, and it says nothing about the content your tools return.

It still points the right way. A harness that keeps a stable, cacheable prefix pays less for every turn that re-sends it. If you migrate, the command palette's "Devin: Open Cascade Migration Wizard" brings workflows and memories over in one pass, according to the same page.

Devin Local also brings subagents. A subagent runs in its own conversation chain, so a noisy subtask doesn't pile its tool output into your main thread.

How do you keep files out of Devin Desktop's context?

Add a .devinignore at the repository root for anything generated, vendored or bulky.

The .devinignore docs say indexing already skips node_modules and hidden paths. The file uses .gitignore syntax, and matched paths "are excluded from indexing and cannot be viewed, edited, or created by Devin". It stacks with .gitignore, so list only what Git still tracks.

This matters because Devin Desktop is retrieval-first. Per its context awareness page, it reads open files and pulls snippets from an index of the whole codebase. Fixtures and snapshots that match a query compete with your real code, and every retrieved chunk is billed as input.

How do rules and memories affect the bill?

Rules ride along with requests, so keep them short and scoped; auto-generated memories don't consume credits.

Devin Desktop now reads rules from .devin/rules/, falls back to .windsurf/rules/, still reads a legacy .windsurfrules, and also picks up AGENTS.md, per the FAQ. Directory rules take an activation trigger in their frontmatter:

• always-on: applies to every request, so it's a fixed cost on each one;
• model-decision: loaded when the model judges it relevant;
• glob: loaded only for matching files;
• manual: loaded only when you invoke it.

Move anything that isn't universal off always-on. A 2,000-character testing guide only needs to load when test files are in play, which is exactly what a glob trigger does.

Cognition's memories and rules page caps rule files at 12,000 characters each and advises rules that are "simple, concise, and specific". It also states that creating and using auto-generated memories does not consume credits. Retrieved memories still occupy context, so prune stale ones.

How do you stop MCP tools from eating context?

Toggle off servers you don't need in this project and list unused tools in disabledTools.

Every advertised tool definition is input on every turn, called or not. For the Cascade agent, the MCP docs set a ceiling of 100 tools at any time. The MCPs section of the Cascade panel's Actions menu shows how many tools each server enables, with a toggle per server.

Devin Local reads its MCP servers from the Devin CLI config files instead of Cascade's mcp_config.json. Whichever agent you use, the habit is the same: one server with forty tools for a job you do weekly is a weekly job billed on every request. The background is in lazy MCP loading.

Which Devin Local modes and commands save tokens?

Plan mode for research, /ask for one-off questions, and /compact or /clear once a session drifts.

Devin Local shares Devin CLI's agent modes, Normal, Plan and Ask, and its essential commands:

Plan mode writes its plan to ~/.devin/plans/plan-<session>.md. That file lets you hand the plan to a fresh session instead of carrying the whole research thread into implementation.

On model choice, Cognition's models page recommends Fusion, which pairs "a frontier lead model with a cost-efficient sidekick". For current prices and quota cost per model, it points you to the selector itself.

What do Devin Desktop's settings leave on the table?

The raw output of the editor's own tools: shell output, web pages and documents still arrive at full size.

Ignore files, scoped rules and lean MCP decide what the agent may read. They don't shrink what a command prints. A failing test suite still lands as thousands of tokens when the failing assertion is twenty, and it's re-sent on every following turn.

That's where Tokenade helps, with one honest caveat. Devin Desktop exposes no hook that can sit in front of its built-in tools. So the compatibility matrix lists its channel as manual: some coverage is automatic, some is yours to invoke.

What does Tokenade do on Devin Desktop?

Two things install automatically; the rest you trigger with three commands.

tokenade install detects the Devin (ex-Windsurf) config folders and then:

• writes its rules to ~/.codeium/windsurf/memories/global_rules.md;
• wraps every third-party server in ~/.codeium/windsurf/mcp_config.json as tokenade mcp-proxy -- <original command>, keeping the original, so MCP results arrive folded. tokenade uninstall restores it.

For the editor's own tools, call Tokenade yourself: tokenade wrap '<cmd>' for commands, tokenade web <url> for pages, tokenade read <file> for documents. The matrix marks command output, web, context and documents Partial+: partial with the binary alone, full once the LLM proxy is set up.

On the open token optimizer benchmark, Tokenade cuts session cost by 39% on long sessions against running no tool at all. The technical brief for agents is at tokenade.net/agents/devin-desktop.md.

How to apply this today

1. Start new work on Devin Local; migrate Cascade memories and workflows with the wizard.
2. Add a .devinignore for fixtures, snapshots, coverage and vendored code.
3. Move rules to .devin/rules/ and switch anything non-universal from always-on to glob or model-decision.
4. Toggle off MCP servers you don't use here; list unused tools in disabledTools.
5. Research in /plan, and /compact or /clear when the task changes.
6. Install Tokenade and route noisy commands through tokenade wrap. The tool-agnostic levers are in how to reduce AI coding agent token usage.

What goes wrong (anti-patterns)

Assuming the rename changed nothing. The editor is the same; the default agent is not. Advice written for Cascade's Flows applies to old conversations, not new tabs.

Making every rule always-on. It feels tidy and bills every request. Scope rules with triggers.

Trusting the 100-tool ceiling as a budget. It's a hard limit on Cascade, not a target. Forty idle tools are forty definitions re-sent per turn.

Counting on compaction to fix bloat. /compact summarises what you already paid for. Keep the noise out first.

FAQ

Is Devin Desktop the same as Windsurf?

Yes. Cognition renamed Windsurf to Devin Desktop on June 2, 2026, through a regular update, with Windsurf settings ported automatically. The editor is the same; the default agent for new conversations is now Devin Local rather than Cascade.

Is Devin Local cheaper than Cascade?

Cognition advertises that most tasks take up to 30% fewer tokens on Devin Local than on Cascade, crediting a greater focus on prompt caching. That's their own comparison between their two agents, not an independent measurement, but it's a reason to run new work on Devin Local.

Where do Devin Desktop rules live now?

In .devin/rules/, with .windsurf/rules/ as a fallback and the legacy .windsurfrules file still read. Devin Desktop also reads AGENTS.md. Global rules stay in ~/.codeium/windsurf/memories/global_rules.md, which is where Tokenade writes its own.

Do Devin Desktop memories cost tokens?

Creating and using auto-generated memories doesn't consume credits, per Cognition's docs. A retrieved memory still takes room in the context window, though, so delete stale ones.

Does Tokenade work with Devin Desktop?

Yes, partly automatically. tokenade install writes a rules file and wraps your MCP servers. Shell output, web pages and documents need tokenade wrap, web or read, because no hook can sit in front of the editor's own tools. The LLM proxy closes most of that gap.

See also:

• How to reduce AI coding agent token usage: the tool-agnostic pillar
• How to reduce Windsurf token usage: the Cascade-era levers and legacy settings
• How to reduce Cursor token usage: the same levers in another AI editor
• Lazy MCP loading: why idle tool manifests still bill