Skip to main content
Tips

beginner tip

Is Claude Code Down, or Is It Your Context?

AIErudit angle: 'the model feels dumber today' is almost always a context-pressure problem you can diagnose in 30 seconds, not a real outage — a core operational skill for anyone shipping with AI. This tip gives a three-step triage (status page, DownDetector, /doctor) and the three context failures (autocompact thrashing, MCP garbage, anchoring) that masquerade as model degradation, plus the /compact and /clear fixes.

claude-codecontext-windowtroubleshootingtips
Tip visual

Published: May 18, 2026 Updated: 2026-06-07

Original: https://wmedia.es/en/tips/claude-code-is-down

TL;DR Three checks in order, 30 seconds each. 1) status.claude.com — Anthropic publishes Claude Code uptime as a separate component (99.14% over the last 90 days) with Atom/RSS, email, SMS, Slack and webhooks. 2) DownDetector — if the status page is slow to reflect an incident, this one collects user reports in real time. 3) /doctor — rules out local issues (broken MCP, ripgrep, invalid settings). If everything's green and Claude still "feels dumber today," the culprit is almost always context pressure (autocompact thrashing), not the model.

There's a recurring meme on Reddit and X: "Claude Code is getting dumber". Sometimes it's real (incident or degradation). Almost always it's one of three perception errors: your context window at 80%, an MCP that dumped 30k tokens of garbage, or anchoring from a too-long conversation. Here's how to tell them apart.

1. status.claude.com — first place to look

Anthropic publishes a status page with per-component uptime. Claude Code is separate from Claude API, Claude.ai, and Claude Cowork:

ComponentWhat it means
Claude CodeThe CLI itself
Claude API (api.anthropic.com)The backend your prompts hit. If this is down, Claude Code is too
Claude.aiThe web chat. Doesn't affect Claude Code
Claude Consoleplatform.claude.com (API key management)

Hit "Subscribe" in the top right to get incidents via email, SMS, Slack, Microsoft Teams, webhook, or Atom/RSS before they hit you.

2. DownDetector — to corroborate

Sometimes Anthropic's status page takes 15-30 minutes to reflect an incident because the team needs to confirm it internally. DownDetector collects user reports in real time:

https://downdetector.com/status/claude-ai/

If DownDetector shows a spike and status.claude.com is still green, wait 10 minutes and check again. Real spike → it's going down even if officially nobody's said so yet.

3. /doctor — rule out local issues

If both checks above come back clean, the problem isn't Anthropic. It's your setup. Run:

/doctor

Audits installation, settings, MCP servers, and context usage in one pass. If Claude Code won't even start so you can't run /doctor:

claude doctor

Same report from outside the CLI. There's a dedicated /doctor tip if you want to dig deeper.

"Claude is dumber today" — what's actually happening

The model doesn't change between sessions. Anthropic ships releases with semver and /release-notes tells you if anything changed in your CLI. What does change is your context. Three symptoms commonly mistaken for model degradation:

Autocompact thrashing. Your context window fills, autocompact reduces it, and the next tool result fills it again. Claude Code stops retrying to avoid burning tokens. Symptom: short, slow responses that feel "lazy". Fix:

/compact "keep only the plan and the diff"
# or, if the earlier conversation isn't needed:
/clear

MCP garbage. An MCP server returns a huge blob (a schema, a DB dump, a Playwright result) and eats half your context. Symptom: the session slows down right after a tool call. Fix: disable that MCP with /mcp or cut the conversation branch with /branch.

Anchoring. You're 200 turns in and the first misread poisoned everything since. Symptom: Claude repeats mistakes even when you point out the right answer. Fix: /clear, start fresh. To preserve what was good, ask Claude to write a /recap first.

If you did all three and nothing changes: it's probably Claude. Report it with /feedback (auto-attaches session context).

Subscribe before the next outage catches you

If your work depends on Claude Code, don't wait to refresh the status page when things feel off:

ChannelBest for
Email / SMSCritical outages, outside the browser
Slack / Microsoft TeamsTeam channel notification
WebhookPlug into your own internal status dashboard
Atom / RSSIf you keep everything in a reader

Subscribe once and you'll know before your team.

Pairs well with

Official docs: Troubleshooting · Status page