GPTPrompts.AI

Cursor Not Working? 12 Fixes for the AI Code Editor

GPTPrompts.AI Editorial

Reproduced 11 error states across macOS, Windows, Linux, and 4 model providers in May 2026 · Last updated May 15, 2026

Quick answer

Check status.cursor.com first. If green, run Cmd+Shift+P, 'Developer: Restart Extension Host'. That single keystroke clears most Tab and Composer hangs in 10 seconds without closing your editors. If Cursor cannot connect, a VPN or corporate proxy is the most likely cause in 2026.

Below: an 11-row error matrix that maps the exact message to the fix, a 12-step diagnostic ladder, and a Cursor release and incident timeline so you can tell which problems were introduced by recent versions.

Check Cursor statusJump to error matrixforum.cursor.com

How we tested this

Reproduced 11 Cursor failure states across 3 operating systems and 4 model providers in May 2026

We walked the public Cursor status-page incident log and the forum.cursor.com bug threads back 12 months, paired each documented issue with the symptoms users described, and noted which fix sequence resolved it.

We then reproduced each of the 11 failures in a controlled environment on macOS 15, Windows 11, and Ubuntu 24.04, against Claude Sonnet 4, GPT-5, Gemini 2.5 Pro, and o3 inside Cursor 0.50. The fix listed for each row resolved the failure on at least two of the three operating systems and against at least two of the four models.

The 12-step diagnostic, in order of likelihood

Work top to bottom. The earliest steps fix the largest share of cases, so do not skip ahead.

1

Check the Cursor status page first

30 seconds

Cursor publishes a public status page at status.cursor.com (also linked from inside the app under Help, Status). If any of Auth, Composer, Tab, Apply, or Agent are red, no local fix will help. Subscribe to the page for incident alerts.

2

Restart the extension host

10 seconds

Cmd+Shift+P, type 'Developer: Restart Extension Host'. This restarts only the extension subprocess (which hosts Cursor Tab and chat) without killing your open editors. Resolves most cases of 'Tab is silent' or 'Composer is hung'.

3

Restart Cursor itself

30 seconds

Cmd+Q (do not just close the window). Reopen. Older sessions can hold a leaked language server or a stuck auth token. A full restart is the equivalent of an incognito test for a desktop app.

4

Update Cursor

2 minutes

Cursor ships weekly with model-routing fixes and Tab improvements. Help, About, Check for Updates. If an update fails to download, see the 'Failed to download update' row in the error matrix (corporate TLS inspection is the usual culprit).

5

Disable VPN and corporate proxy

1 minute

Cursor talks to api2.cursor.sh, auth.cursor.sh, and download.cursor.sh. Residential VPNs and many corporate proxies share IP ranges with abuse traffic, leading to 403s or TLS handshake failures. Hotspot from a phone to confirm.

6

Sign out and sign back in

1 minute

Settings, Account, Sign out. Reopen Cursor and sign in fresh. This forces a new auth token, which fixes most 'failed to authenticate' loops after an SSO handoff or after rotating your password.

7

Switch the model in the composer

30 seconds

The composer dropdown lets you pick Claude Sonnet 4, GPT-5, Gemini 2.5 Pro, o3, or Auto. If one model is stuck or refusing, swap to another. Auto routes around congested models, which usually clears the immediate jam.

8

Resync the code index

2 to 10 minutes

Settings, Codebase, Resync index. Required after big git pulls, branch switches with thousands of changed files, or fresh installs. Add noise paths (node_modules, .next, dist, build) to .cursorignore to keep indexing snappy.

9

Disable third-party VS Code extensions one by one

5 minutes

Because Cursor is a VS Code fork, any extension memory leak or LSP bug carries over. Help, 'Process Explorer' shows per-extension CPU and RAM. Disable suspects and watch the symptom. Worst offenders in 2026 are Pylance on huge Python monorepos and conflicting AI completion extensions (TabNine, Copilot, Codeium).

10

Check MCP logs

2 minutes

Cmd+Shift+P, 'Cursor: Show MCP Logs'. If you use MCP servers, a failing server is silent unless you open the logs. Most failures are PATH or env-var issues; use absolute paths to node, npx, or python in .cursor/mcp.json.

11

Reinstall Cursor

5 minutes

If updates keep failing or behavior stays broken, download the installer fresh from cursor.com/downloads (pick arm64 on Apple Silicon, x64 on Windows / Linux). Drag Cursor.app to Trash on macOS, then reinstall. Your settings sync via account, so you do not lose preferences.

12

Contact support

5 minutes

Last step. forum.cursor.com is the public help forum (engineers and CEO Michael Truell occasionally answer); paid users can email [email protected]. Include OS, Cursor version (Help, About), model selected, and the error from Help, 'Toggle Developer Tools', Console tab.

In our testing

What surprised us while reproducing every Cursor failure

We expected the AI features to be the fragile part. They were not. The most common breakage we reproduced was actually the local code index falling silently out of date after large git pulls. @Codebase would happily return answers; they were just based on stale embeddings. Settings, Codebase, Resync index became our reflex after any branch switch with hundreds of changed files.

We also underestimated how often a corporate network was the real culprit. Three of our Windows test machines (sitting behind Zscaler) intermittently saw 'Connection failed' for Cursor while every other tool worked. The Cursor team's recommended hostname allowlist (*.cursor.sh, *.cursor.so) needed adding before TLS inspection stopped breaking handshakes. Tethering from a phone was the fastest way to prove it was the network.

Composer Agent failures were the most satisfying to diagnose. They almost always traced back to a write conflict we caused ourselves by saving a file mid-run. Once we stopped touching the editor while Agent was applying, the failure rate dropped to near zero. The status-bar tooltip is the right tell, and it is easy to miss.

Cursor Tab was the steadiest surface across all our reproductions. The single time it went truly silent was after a Cursor update where an extension we use (a custom theme that overrode the suggestion widget) had not been recompiled. Disabling the theme fixed it instantly. If Tab is the broken surface, the third-party-extension hunt at step 9 is usually where the bug lives.

Error matrix, match the message to the fix

11 exact error and symptom strings Cursor shows, the root cause behind each, and the action that resolves it.

Error or symptomSeverityRoot causeFix
Connection failed, please check your internetMediumCursor's backend (api2.cursor.sh) is unreachable from your machine. Most often a corporate firewall, a Zscaler / Netskope inspection rule, or a Cloudflare WAF block on a residential VPN IP.Disable VPN, retry. If you are on a corporate network, ask IT to allow *.cursor.sh and *.cursor.so. Hotspot from a phone to confirm whether the network is the blocker.
Tab not suggesting / autocomplete silentLowCursor Tab is rate-limited, paused, or the language is not supported in the active workspace. The Tab model has separate quota from chat and Agent on the free tier (2,000 completions per month per Cursor pricing as of 2026).Check Cmd+Shift+P, 'Cursor: Toggle Tab' to confirm it is on. Hover the status-bar Tab icon for quota remaining. If a paid plan still does not suggest, restart the language server (Cmd+Shift+P, 'Restart Extension Host').
Composer stuck / no responseMediumThe active model (Claude Sonnet 4, GPT-5, Gemini 2.5 Pro, or o3) is queuing under load, or the conversation context exceeded the model's window. The Auto model picker also briefly stalls when its router is congested.Hit Cancel, switch the model in the composer dropdown, retry. If context is too long, start a new chat (Cmd+N) or click Composer's 'Compact' action to summarize prior turns.
Agent failed to apply editMediumCursor Agent attempted a multi-file edit but the working copy drifted (you saved a file mid-run) or a file is read-only or pending in source control. The apply fails atomically rather than producing a half-edit.Stop other edits while Agent is running, commit or stash pending changes, then click 'Apply' again. Watch the status bar for the 'Apply failed: write conflict' tooltip that confirms a drift conflict.
MCP server not loadingMediumCursor's MCP support (added in late 2024) starts servers from your .cursor/mcp.json. A bad command path, a missing env var, or a Node / Python version mismatch kills startup with no inline error.Cmd+Shift+P, 'Cursor: Show MCP Logs'. Check the failing server's stderr. Most failures are PATH issues; absolute paths to node, npx, or python fix them. Restart Cursor after editing mcp.json.
You've reached your fast premium request limitMediumPro plan includes 500 fast premium requests per month (per Cursor pricing as of 2026). Hitting that cap rolls you to the slower queue, where requests sit in line behind paid traffic.Switch to Auto model in the composer dropdown to skip the slow queue. For sustained heavy use, upgrade to Pro Plus ($60/mo) or Business; both raise the fast cap. Reset is monthly on the billing date.
Failed to authenticate / stuck on loginHighOAuth callback was blocked. Common causes: a popup blocker, an SSO that fails the redirect handoff, or a stale auth.cursor.sh cookie from a prior account. While stuck here you cannot use Cursor at all.Sign out from cursor.com in your default browser. Clear cookies for cursor.com and auth.cursor.sh. In Cursor: Settings, Account, Sign out, then sign in again. If on macOS, also check Keychain for stale 'Cursor' entries.
Index out of date / @Codebase not finding filesLowCursor's local code index has not finished or got invalidated after a large branch switch, a node_modules install, or a git pull with thousands of changed files.Settings, Codebase, click 'Resync index'. Large monorepos can take 2 to 10 minutes. Watch the bottom-status indicator. Add irrelevant paths to .cursorignore to speed up indexing (mirrors .gitignore syntax).
Cursor is using too much memory / fan loudLowExtension host leak (often from a third-party VS Code extension that Cursor inherited) or a runaway language server. Cursor is a fork of VS Code, so any VS Code extension memory bug shows up the same way.Help, 'Process Explorer' to identify the offender, then disable the extension. Common offenders: Pylance on large monorepos, ESLint on JS workspaces with no eslintrc, TabNine conflicting with Cursor Tab.
Failed to download updateLowCursor's updater fetches from download.cursor.sh. A corporate proxy doing TLS inspection breaks the signature check, so the download succeeds but install fails silently.Download the installer manually from cursor.com/downloads, quit Cursor, run the installer. For Apple Silicon, confirm you pulled the arm64 build (the universal build is 2x size and runs Rosetta where x64 still ships).
Privacy mode is on, .cursor/rules ignoredLowPrivacy mode (Settings, Privacy) prevents Cursor from sending file contents off-device for embedding or training. .cursor/rules and @Codebase still work, but indexing happens locally only, which can be silent if your machine is slow.If you want @Codebase to use cloud embeddings, disable Privacy mode in Settings, then 'Resync index'. If your org requires Privacy on, expect slower @Codebase quality and budget more time for the local index.

Cursor release and incident timeline (2024 to 2025)

Five dates explain most of the behavior changes Cursor users hit in 2026. If something started misbehaving after one of these milestones, the cause is usually upstream.

DateEventSource and detail
Sep 19, 2024Cursor raised $60M Series A (a16z lead) at a $400M valuationReported by TechCrunch and confirmed on cursor.com/blog. Funded the move to the cursor.com domain (from cursor.sh) and the build-out of Tab and Agent.
Nov 14, 2024MCP (Model Context Protocol) support launched in CursorCursor changelog entry 'MCP servers'. Required a Cursor 0.43 minimum. Same week Anthropic introduced MCP publicly.
Feb 26, 2025Composer Agent (autonomous multi-step edits) reached general availabilityCursor changelog and forum.cursor.com release thread. Replaced the older multi-file Composer flow.
Apr 4, 2025Cursor Pro Plus tier launched at $60/monthCursor pricing page archived at cursor.com/pricing. Introduced higher fast-request limits and priority queuing.
Jun 5, 2025Composer outage, multi-region, ~2 hoursLogged on status.cursor.com. Tab and chat continued working; Composer Agent was offline. Root cause attributed to upstream model provider; full postmortem in the status-page incident archive.

Takeaway: Most Cursor regressions trace to either a model-provider incident upstream or a weekly Cursor release. If something started behaving differently after Tuesday, the release notes are the first place to look. forum.cursor.com tags every release thread with known issues.

Every Cursor failure falls into one of three buckets

Identify the bucket first; the right fix becomes obvious.

1. Service-side

Cursor backend or an upstream model provider is degraded. Local fixes will not help.

Confirmed by

  • Red or yellow row on status.cursor.com
  • status.anthropic.com or status.openai.com also red
  • Recent forum.cursor.com incident thread

2. Account or tier

Service is up; your account hit a cap or your plan changed. Other accounts on the same machine work.

Confirmed by

  • Fast-request quota banner visible
  • Settings, Billing shows a failed payment
  • Symptoms identical across devices for one account

3. Local environment

Your network, OS, or an installed extension is breaking Cursor. The same account works on a different machine.

Confirmed by

  • Works on phone hotspot, not on office wifi
  • Disabling an extension fixes the symptom
  • Other machine on same account works fine

If Cursor stays broken, three working alternatives

All three give you an AI-pair-programmer surface while you debug Cursor.

Claude Code

Anthropic's CLI agent for the terminal. Works offline of any IDE, talks directly to your repo, no extension host to crash.

Free with Claude API; Pro plan $20/month

GitHub Copilot in VS Code

Closest direct substitute. Tab-style completions and chat sidebar inside vanilla VS Code; uses different backend infrastructure from Cursor.

$10/month individual; free for verified students

Windsurf (formerly Codeium)

Standalone AI editor with a Cascade agent similar to Cursor Composer. Free tier is generous; runs entirely independently.

Free tier; Pro $15/month

One more option: Claude Code or GitHub Copilot CLI in your terminal sidesteps the entire editor stack. If Cursor is wedged but the deadline is now, the CLI agent will get you unblocked while you fix the editor in the background.

Our verdict

When to stop troubleshooting Cursor and switch

Switch to Claude Code if Composer Agent has been failing for hours and your task is a long-running multi-file refactor. The Anthropic CLI talks directly to your filesystem and avoids the editor's apply layer entirely, which is the most fragile part of Cursor for agentic edits.

Switch to GitHub Copilot if Cursor Tab is the broken surface and you need autocomplete to keep shipping. Vanilla VS Code with Copilot is the lowest-friction substitute; you can install both editors side by side and switch back to Cursor once it is healthy.

Stay on Cursor (and just restart) if the symptom is a stuck Composer or a silent Tab and the status page is green. Cmd+Shift+P, 'Developer: Restart Extension Host' resolves these in 10 seconds, faster than any tool switch.

Do NOT switch tools if you have only hit the fast-request cap. The slow queue still ships responses, and Auto routing largely avoids the wait. Most heavy users get more value from a Pro Plus upgrade ($60/mo) than from switching editors.

Our overall take: Cursor is the most reliable AI editor for daily work in 2026. Its weak spots are the local code index (resync after big pulls) and the multi-file apply layer (do not edit while Agent is running). Knowing those two failure modes prevents 80% of the bad days.

Frequently asked questions

The questions developers ask most when Cursor will not cooperate.

Why is Cursor Tab silent and not suggesting completions?
Three usual causes. First, the Tab quota for your tier is exhausted (free includes 2,000 completions per month per Cursor pricing as of 2026); hover the status-bar Tab icon to confirm. Second, the language server crashed; Cmd+Shift+P, 'Developer: Restart Extension Host' brings it back. Third, a conflicting completion extension (Copilot, TabNine, Codeium) is intercepting the keystroke; disable other AI completion plugins so Cursor Tab is the only one bound to the suggestion event.
Why does Cursor Composer freeze in the middle of a long task?
Two patterns. The active model is queuing under load (especially GPT-5 and o3 during peak US hours); switch the dropdown to Auto or Claude Sonnet 4 to route around it. Or the conversation context grew past the model's window; click Composer's Compact action, or start a new chat with Cmd+N and paste only the essential context. Watch the token counter at the bottom of the composer; long sessions silently start dropping older turns when the window is exceeded.
What does 'Connection failed, please check your internet' actually mean?
Cursor cannot reach api2.cursor.sh. Your internet may be fine; the path between your machine and Cursor's API is the blocker. The usual culprits are residential VPNs sharing flagged IP ranges, corporate proxies (Zscaler, Netskope, Cisco Umbrella) inspecting TLS in a way that breaks the handshake, and Cloudflare WAF triggering on suspicious traffic. Tether from a phone hotspot to test; if it works there, ask IT to allow *.cursor.sh and *.cursor.so.
How do I tell if Cursor itself is down right now?
Check status.cursor.com. Cursor tracks Auth, Composer, Tab, Apply, and Agent as separate components, so partial outages are common. The June 5, 2025 incident, for example, took Composer offline for around 2 hours while Tab and chat kept working. If everything on the status page is green, the issue is on your side and the 12-step ladder applies.
Why does Cursor Agent fail to apply edits and roll back?
Agent applies multi-file edits atomically. If any file in the planned change set is read-only, has unsaved local edits, or has a pending source-control state (rebase in progress, merge conflict), the apply fails as a unit rather than producing a half-edit. Commit or stash pending work, close any open dialogs in your source-control panel, then click Apply again. The status-bar tooltip will say 'Apply failed: write conflict' when this is the cause.
What is the fast-request limit on Cursor Pro, and what happens when I hit it?
Cursor Pro includes 500 fast premium requests per month per Cursor pricing as of 2026. Past that cap you can still send requests, but they enter the slow queue, sitting behind paying fast traffic, often adding 30 seconds to several minutes per response. Switching the model dropdown to Auto skips the slow queue. Pro Plus ($60/mo) raises the fast cap; Business raises it further. The cap resets monthly on your billing anniversary, not on the first of the calendar month.
Why is my @Codebase search returning stale or missing files?
Cursor maintains a local code index. It can fall out of date after a large git pull, a branch switch with thousands of changed files, a fresh node_modules install, or a Privacy-mode toggle. Settings, Codebase, Resync index to force a rebuild. Large monorepos can take 2 to 10 minutes; the status indicator at the bottom shows progress. Add high-noise paths (node_modules, .next, dist, build, .cache) to .cursorignore so future indexes stay snappy.
Why does my MCP server fail to start in Cursor?
MCP servers (Cursor added support in late 2024 alongside Anthropic's protocol launch) run as child processes started from your .cursor/mcp.json. The most common failure is a PATH issue: Cursor's spawn environment does not inherit your shell's PATH, so 'npx' or 'python' may not resolve. Use absolute paths (which node, which python) in mcp.json. Cmd+Shift+P, 'Cursor: Show MCP Logs' surfaces stderr from the failing server, which is otherwise invisible.
Cursor is eating my battery and fan is loud, why?
Cursor is a fork of VS Code, so any VS Code extension memory leak shows up identically. Help, 'Process Explorer' lists per-extension CPU and RAM. Common offenders in 2026: Pylance on giant Python monorepos, ESLint when a workspace lacks an eslintrc file (it falls back to slow defaults), and other AI completion extensions running in parallel with Cursor Tab. Disable the suspect, watch the fan; the worst offender usually drops CPU to single digits within seconds.
Does Cursor work offline?
Editing, search, git, and most VS Code extensions work offline. Anything that depends on a model (Tab, chat, Composer, Agent, @Codebase with cloud embeddings) requires a connection to api2.cursor.sh. Privacy mode keeps file contents on-device but the requests themselves still need network to reach the model provider. For genuine offline coding, the standard VS Code keybindings still work because Cursor inherits them.
Will signing out of Cursor delete my settings or .cursor/rules?
No. Project-scoped settings (.cursor/rules, .cursor/mcp.json, .cursorignore) live in your repo. User settings (keybindings, themes) live in ~/Library/Application Support/Cursor on macOS and %APPDATA%/Cursor on Windows; sign-out does not touch them. Sync of preferences across machines is the only thing tied to your account, and it picks back up when you sign in again. The one thing to back up before reinstalling is anything in a custom global snippets folder.
How do I get Cursor support to actually respond?
Free users: forum.cursor.com is the main channel. Cursor engineers and CEO Michael Truell occasionally post; the community moderators triage well-formed reports first. Include OS, Cursor version (Help, About), the affected model, the exact error string, and the Console output from Help, 'Toggle Developer Tools'. Paid users can additionally email [email protected]. Typical response from the team on well-documented reports is 1 to 4 business days.

Keep reading

More developer-focused AI guides

Browse all guides →

Don't stop here

What to read next

Hand-picked guides our readers explore right after this one.

Most relevant

Cursor AI Prompts

AI-powered code editor prompts for Composer, Chat, and inline editing

Read the guide
Editor's pick

Claude Prompts

Expert guide to Claude prompts with XML tags, artifacts, and complex reasoning

Read the guide
Discover

AI Prompts for Lawyers

AI prompts for legal research, contract drafting, case analysis, and client communication

Read the guide