deadends.dev

deadends.dev

Stop AI agents from repeating known failures - in code AND in the real world.

AI assistants reliably fumble two kinds of problems: known-failed code fixes, and country-specific real-world rules they've never been exposed to in training. deadends.dev now covers both:

• Code errors (2,089 entries, 51 domains): what NOT to try when an agent hits ModuleNotFoundError, CUDA OOM, CrashLoopBackOff, etc.
• Country-scoped dead ends (250+ entries across 52 countries): visa rules (ETA/eVisitor, NZeTA, e-visas, arrival cards), banking requirements, legal red lines (lèse-majesté, §86a, Article 301), cultural taboos (chopsticks in rice, clock gifts in China, red-ink names in Korea), food safety (tap-water safety by country), emergency numbers, driving norms (left-hand traffic), housing contracts - all the friction where a plausible-sounding global answer is wrong locally.

Why the expansion? Coding dead ends are largely solved by a good LLM.

Country-specific friction - Japanese hanko requirements, Schengen 90/180

math, Ramadan business hours, Saudi alcohol ban, Indian beef taboos - is

where generic AI advice breaks hardest. The codebase and schema are

identical; the env segment just carries a country code.

90% Precision@1 · 0.935 MRR · Data Quality Dashboard

Website: deadends.dev · MCP Server: Smithery · PyPI: deadends-dev · API: /api/v1/index.json

Repository: https://github.com/dbwls99706/deadends.dev

Why Use This?

What makes this different from asking an LLM?

• Deterministic: Same query → same answer, every time. No hallucination.
• Country-scoped: ID format {domain}/{slug}/{env} - env holds the country code (kr, jp, us, de...) so the same taboo can be answered differently for different jurisdictions.
• Primary-sourced: Every country canon cites government sites, embassies, or verifiable reporting. No "based on general knowledge" answers.
• Community-validated: Fix success rates updated from real outcome reports.
• Sub-millisecond: Local regex matching, no API roundtrip.

현실적인 한계 (운영 관점)

• 모든 에러를 다 커버하지는 못합니다. 없는 케이스는 이슈/PR/report_outcome로 빠르게 보완합니다.
• 설명의 깊이보다 실전 해결 우선(dead end/workaround 중심)으로 설계되어 있습니다.
• 신뢰성은 도메인/케이스마다 다를 수 있으므로, 고위험 변경은 공식 문서/벤더 가이드와 교차 검증을 권장합니다.

Quick Start (30 seconds)

pip install deadends-dev

deadends "CUDA error: out of memory"

MCP Server (Claude Desktop / Cursor)

Add to ~/.claude/claude_desktop_config.json:

{

"mcpServers": { "deadend": { "command": "python", "args": ["-m", "mcp.server"], "cwd": "/path/to/deadends.dev" } } }

Or install via Smithery (no local setup):

npx -y @smithery/cli@latest install deadend/deadends-dev --client claude

MCP Unauthorized 빠른 해결 가이드 (사람용)

deadend: calling "initialize": sending "initialize": Unauthorized 에러가 보이면 아래를 순서대로 그대로 실행/확인하세요.

1. 로컬 서버 모드인지, 원격(Smithery) 모드인지 하나만 사용

# 로컬 서버 확인 (정상 시 툴 목록이 출력됨)

python -m mcp.server --help

1. Claude Desktop 설정 파일 점검 (cwd는 실제 경로여야 함)

cat ~/.claude/claude_desktop_config.json

1. 로컬 서버 직접 실행 테스트

cd /path/to/deadends.dev

python -m mcp.server

1. Smithery 모드라면 재설치(토큰/설정 꼬임 복구)

npx -y @smithery/cli@latest uninstall deadend/deadends-dev --client claude

npx -y @smithery/cli@latest install deadend/deadends-dev --client claude

1. 마지막으로 Claude Desktop 완전 재시작

# macOS 예시

osascript -e 'quit app "Claude"' open -a Claude

팁: Unauthorized는 보통 잘못된 cwd, 중복 서버 설정(로컬+원격 동시), 또는 만료된 인증 상태에서 발생합니다.

Antigravity (Google AI IDE)

Add as a remote MCP server - no authentication required:

{

"mcpServers": { "deadend": { "serverUrl": "https://deadends.dev/mcp", "type": "http" } } }

Note: Antigravity uses serverUrl (not url). If you get Unauthorized, remove any existing deadend entries from the MCP Store and re-add manually using the config above. See the Antigravity MCP auth guide for general troubleshooting.

Python SDK

from generator.lookup import lookup, batch_lookup, search

Single error lookup

result = lookup("ModuleNotFoundError: No module named 'torch'")

What NOT to try (saves tokens and time)

for d in result["dead_ends"]: print(f"AVOID: {d['action']} - fails {int(d['fail_rate']*100)}%")

What actually works

for w in result["workarounds"]: print(f"TRY: {w['action']} - works {int(w['success_rate']*100)}%")

Batch lookup (multiple errors at once)

results = batch_lookup(["error1", "error2", "error3"])

Example Response

## ModuleNotFoundError: No module named 'X' (Python 3.11+)

Resolvable: true | Fix rate: 0.88

Dead Ends (DO NOT TRY):

• pip install X with system Python (fails 70%): venv not activated

Workarounds (TRY THESE):

• Create venv, activate, then pip install (works 95%)
• Use python -m pip install instead of bare pip (works 90%)

MCP Tools (11)

API Endpoints

Covered Domains (54)

Code error domains (51)

Country-scoped real-world domains (new, growing)

Data Quality

All metrics are publicly available on the Data Quality Dashboard:

• 2,204 canon entries across 54 domains and 39+ countries
• Benchmark: 90% Precision@1, 95% Precision@3, 0.935 MRR (on code scenarios)
• Error transition graph: 4,330+ edges connecti