Cursor AI Error: 7 Real Fixes That Actually Work (2026)

Error	Quick Fix	Time
“Connection failed” / check your internet or VPN	Turn off VPN/proxy, or set the correct proxy in Cursor’s network settings, then reload	2 min
Stuck on “Setting up indexing…”	Cancel indexing, delete the codebase index, add a .cursorignore, re-index	5-10 min
Sign-in loop or [unauthenticated] error	Sign out, clear the local auth cache, sign in again through the browser flow	3 min
“You’ve hit your rate limit” / usage limit reached	Switch to a non-premium/included model or wait for the reset window; check the usage dashboard	1 min
Agent or tool call fails silently (MCP)	Restart the MCP server from Settings > MCP, check the Output panel for the real error	5 min
Tab completions slow or hanging	Disable conflicting extensions, shrink open-file context, restart the language server	5 min
“Agent execution timed out”	Break the task into smaller steps, reload the window, retry	2 min

You’re mid-refactor, the agent is halfway through a change, and then it just stops — “Connection failed,” or the spinner on “Setting up indexing…” that never finishes, or a rate-limit banner even though you swear you just started your session. It’s one of the more annoying failure modes in software: the tool isn’t crashing, it’s just quietly not doing the thing you need it to do, and the error message rarely tells you why.

The good news is that almost none of this is mysterious once you know where to look. Cursor is built on VS Code, so a lot of its plumbing — networking, extension host, logs — behaves exactly like VS Code under the hood, and most of these errors trace back to one of five things: a backend outage, a proxy/firewall blocking the AI request, a stale local index or auth token, an extension conflict, or you genuinely hitting a usage cap. Below is the actual troubleshooting path, pulled from Cursor’s own forum threads and status history, not generic “restart your computer” advice.

What Causes Cursor Errors

Backend load and outages. Cursor routes every AI request through its own infrastructure before it ever reaches a model provider (OpenAI, Anthropic, Google, xAI). When that layer is under heavy load or a provider is degraded, you get generic failures like “Connection failed” even though your own internet is fine — this has happened repeatedly through 2026, including widely-reported high-load incidents in mid-2026 that hit users worldwide at once.

Network interference from proxies, VPNs, and corporate firewalls. Cursor’s chat and agent traffic goes out over HTTPS to api2.cursor.sh and related endpoints. Split-tunnel VPNs, SSL-inspecting corporate proxies, and some antivirus/firewall products break or silently strip these connections, which is why “check your internet connection or VPN” is one of the most common error dialogs on the forum.

A corrupted or half-built codebase index. Cursor indexes your project locally (embeddings + a Merkle tree of file hashes) so chat and agent can reference your code. Large repos, symlink loops, huge generated folders (node_modules, build output, vendored code) without a .cursorignore, or an interrupted sync can all leave the index stuck at “Setting up indexing…” indefinitely.

Stale authentication tokens. Cursor sign-in happens through a browser-based OAuth flow, and the desktop app then holds a session token locally. That token can go stale after a password reset, an SSO change, switching networks, or just sitting idle overnight — enterprise users in particular have reported recurring morning re-auth failures.

Usage limits and rate limits. Since Cursor moved to usage-based pricing, “included” requests on a plan are metered, and premium models draw down faster than others. Once you’re over the included allotment (or hit a burst rate limit), requests get throttled or blocked until the next reset or until you enable pay-as-you-go overage.

Quick Fix — Try This First

Before anything else: reload the window. Open the Command Palette (Ctrl/Cmd+Shift+P), run Developer: Reload Window, and let Cursor re-establish its connections from scratch. This sounds too simple to matter, but because Cursor inherits VS Code’s extension host architecture, a huge share of “stuck,” “not responding,” and transient connection errors are really a hung background process — the language server, the indexing worker, or the chat websocket — that a reload just kills and restarts cleanly. It fixes more than a full app restart does, and it’s faster.

If a reload doesn’t clear it, check status.cursor.com before you touch a single setting. If Cursor’s backend is degraded, no local fix will help — you’re just waiting it out.

Step-by-Step Fix Guide

Step 1: Rule Out a Cursor-Side Outage

Check status.cursor.com and the forum’s Bug Reports category. Cursor has had several documented multi-hour incidents in 2026 where AI requests failed globally across local, IDE, and Agent modes simultaneously — if you see a fresh wave of “Connection failed” or “Internal Server Error” threads posted within the last hour, it’s not your setup. Switching models or restarting won’t help; you just have to wait for their status page to go green.

Step 2: Fix “Connection Failed” From Your Side

If status.cursor.com is clean, the problem is almost always on your network path:

Temporarily disable your VPN and retry. If it works, the VPN’s routing or DNS is the culprit — try a different server/protocol, or add an exception for Cursor’s domains.
On a corporate network, check whether an SSL-inspecting proxy is intercepting HTTPS traffic. This breaks Cursor’s certificate validation and produces exactly this error. You (or IT) need to install the corporate root CA where Cursor/VS Code trusts it, or add Cursor’s endpoints to the proxy’s bypass list.
Set an explicit proxy in Cursor if your network requires one. Go to Settings > General > Network, or set it directly in settings.json:

{
  "http.proxy": "http://proxy.company.com:8080",
  "http.proxyStrictSSL": false
}

Temporarily disable antivirus/firewall software that does its own HTTPS scanning (some consumer AV suites are notorious for this) and see if the error clears.

Step 3: Re-authenticate for Login and [unauthenticated] Errors

If chat or agent throws an authentication error, or you’re stuck in a sign-in loop between the desktop app and the browser: sign out completely from Cursor’s account menu, close the app, then sign back in and let the browser-based flow finish before switching back to the editor. If it loops, try signing in from a fresh browser profile (extensions or cached cookies on your default profile can interfere) — enterprise/SSO users specifically have reported this resolving repeated overnight auth failures.

Step 4: Clear and Rebuild a Stuck Index

When indexing hangs on “Setting up indexing…”:

Open the Codebase Index panel (Settings > Indexing & Docs) and cancel the current index.
Add a .cursorignore file at your project root excluding large, irrelevant directories:

node_modules/
dist/
build/
.next/
*.lock
vendor/
coverage/

Delete the local index for that project and re-index. On most setups this is triggered from the same Indexing panel with a “Resync Index” or “Delete and Rebuild” action.
If it still hangs, check for symlink loops or an unusually large monorepo — indexing chokes on both. Point Cursor at a subfolder as a workaround if the full repo is the issue.

Step 5: Check Model Selection and Usage Limits

A lot of “it just stopped working” reports turn out to be rate limits, not bugs. Open Settings > Usage to see your current consumption against your plan’s included allotment. If you’re on a fixed plan and maxed out on a premium model, either switch to an included/cheaper model in the model picker, wait for the reset window shown on that page, or turn on usage-based pricing for overage if your plan supports it. If you’re using your own API key, double-check that key’s provider-side rate limit too — Cursor will surface the provider’s error, which looks identical to Cursor’s own limit at a glance.

Step 6: Fix Slow or Hanging Tab Completions

If completions lag several seconds or never appear: disable other AI-adjacent extensions (GitHub Copilot, other completion providers) since two suggestion engines fighting for the same trigger is a common cause of both slowness and dropped suggestions. Also try trimming how much open-tab context Cursor pulls in — a huge number of open files or an enormous single file both slow down Tab noticeably. Restarting the TypeScript/language server for your project (Command Palette > “Restart TS Server” or the equivalent for your language) often clears completion stalls without a full reload.

Step 7: Update Cursor

Cursor ships frequent point releases, and a chunk of forum bug reports turn out to already be fixed in the next build. Check Cursor > Check for Updates (or Help menu on Windows/Linux). If you’re several versions behind, update and retest before digging further — it costs you a minute and rules out a whole category of already-patched bugs.

Advanced Fixes

Read the actual logs before guessing. Open the Output panel (View > Output) and switch the dropdown to “Cursor” or “Cursor Agent” — the real error (a specific HTTP status code, a timeout value, a malformed MCP response) is almost always more specific than the toast notification you see in the UI. For deeper issues, Cursor also writes logs to disk:

macOS: ~/Library/Application Support/Cursor/logs/
Windows: %APPDATA%\Cursor\logs\
Linux: ~/.config/Cursor/logs/

MCP server failures. If an agent tool call fails or an MCP-connected tool silently stops responding, check Settings > MCP for a red/error status next to the server. Toggle it off and on to force a restart, and confirm the server process itself is still running if it’s a local stdio server (a crashed local process is a very common cause of “tool call failed” that has nothing to do with Cursor itself).

Corporate SSL inspection. If your org’s proxy does TLS interception, you may need to point Cursor/Node at your custom CA bundle explicitly rather than disabling strict SSL (which is a real security downgrade — only use proxyStrictSSL: false temporarily to confirm this is the cause, then switch to the certificate approach):

{
  "http.proxy": "http://proxy.company.com:8080",
  "http.proxySupport": "on",
  "http.systemCertificates": true
}

and set the NODE_EXTRA_CA_CERTS environment variable to your corporate root CA .pem file before launching Cursor.

Reinstall vs. clean profile. A full reinstall rarely fixes anything a clean profile won’t, and it’s more disruptive. Try renaming Cursor’s settings/state folder first so it regenerates fresh — this preserves your actual application binary and just resets local state:

macOS: rename ~/Library/Application Support/Cursor
Windows: rename %APPDATA%\Cursor
Linux: rename ~/.config/Cursor

Relaunch, sign in again, and see if the problem follows you. If it doesn’t, something in your old profile (a corrupted setting, a bad extension state, stale auth cache) was the cause, and you can selectively bring back your keybindings/settings.json rather than starting completely from zero. Only fall back to a full uninstall/reinstall if a clean profile doesn’t resolve it — Cursor’s own forum has multiple threads confirming reinstall alone often doesn’t fix connection issues that are actually network or account-side.

Still Not Working? Try These Instead

GitHub Copilot — the most mature completions/chat integration for VS Code itself, worth a shot if the issue is specifically Cursor’s own infrastructure rather than AI coding tools in general.

Windsurf — Cognition’s AI-native editor, also a VS Code fork, with its own agent (Cascade) and a comparable indexing/agent workflow if you want a near-drop-in alternative while Cursor sorts out an outage.

Claude Code — Anthropic’s terminal-and-IDE-integrated agent; it doesn’t depend on Cursor’s infrastructure at all, so it’s a solid fallback specifically when the problem is Cursor’s backend being down rather than your local setup.

Cline — an open-source, bring-your-own-API-key agent extension that runs inside VS Code (and Cursor itself); useful when you want to isolate whether an issue is “AI agents in general” or specifically Cursor’s own service layer.

FAQ

Is Cursor down right now?

Check status.cursor.com first — it’s Cursor’s official status page and reflects real incidents, not just individual bug reports. If it’s clean but you’re still stuck, cross-check the Bug Reports category on forum.cursor.com; a sudden cluster of identical fresh threads is a reliable signal of an unreported or in-progress incident even before the status page updates.

Why does Cursor say “Connection failed” even though my internet is fine?

This error means Cursor’s app couldn’t complete a request to its backend, which isn’t the same as your internet being down. The usual causes are a VPN or proxy interfering with the HTTPS connection, a corporate firewall doing SSL inspection without the right certificate trust, or Cursor’s own servers being under load. Work through Step 2 above before assuming it’s your ISP.

Does reinstalling Cursor delete my settings and chat history?

A standard reinstall usually preserves your settings.json and local state, but it’s not guaranteed across every OS and version, and it won’t fix a problem caused by your account or network. Rename (don’t delete) your Cursor config folder first, as described in Advanced Fixes, so you can recover it if a clean profile doesn’t actually solve anything.

I’m on a paid plan — why do I keep hitting rate limits?

Since Cursor moved to usage-based pricing, paid plans include a metered amount of usage rather than truly unlimited requests, and premium models consume that allotment faster than lighter ones. Check Settings > Usage to see exactly what’s being consumed and by which model, and either switch models, wait for the reset, or enable pay-as-you-go overage if it’s available on your plan.

Why does the agent say a tool call failed but give no real detail?

This is almost always either an MCP server that crashed or lost connection, or the model itself producing a malformed tool call under load. Check the Output panel’s Cursor Agent log for the actual error rather than the toast message, and restart the specific MCP server from Settings > MCP — that resolves the majority of these reports on the forum.