deadends.dev is structured failure knowledge infrastructure for AI agents, with the core goal of stopping AI agents from repeating known failures—including both code errors and real-world country-specific rules.
Purpose
AI assistants often make two types of mistakes: first, trying known-ineffective code fixes, and second, giving seemingly reasonable but globally generic answers that do not comply with local regulations or culture. deadends.dev provides a structured knowledge base for these two types of problems, helping agents avoid repeating failures.
Main Capabilities
- Code Error Knowledge Base: Contains 2,089 entries covering 51 domains, recording fixes that agents should not try when encountering errors such as ModuleNotFoundError, CUDA OOM, CrashLoopBackOff, and others.
- Country-Scoped Real-World Rules: Contains 250+ entries covering 52 countries, involving visa rules (ETA/eVisitor, NZeTA, e-visa, arrival cards), banking requirements, legal red lines (such as lèse-majesté, §86a, Article 301), cultural taboos (such as sticking chopsticks into rice, giving clocks in China, writing names in red pen in Korea), food safety (tap water safety by country), emergency phone numbers, driving norms (left-hand traffic), housing contracts, and more.
- Deterministic Queries: The same query returns the same answer, with no hallucination.
- Country-Scoped Qualification: IDs are in the format {domain}/{slug}/{env}, where env carries the country code (kr, jp, us, de, etc.), so the same taboo can give different answers for different jurisdictions.
- Primary Sources: Every country entry cites government websites, embassies, or verifiable reporting, with no “based on general knowledge” answers.
- Community Validation: Fix success rates are updated based on real outcome reports.
- Sub-millisecond Response: Local regex matching, with no API round trip.
- Error Transition Graph: 4,330+ edges connecting related errors.
- Data Quality Dashboard: Publicly discloses all metrics, including 2,204 canonical entries, 54 domains, and 39+ countries.
How to Use
- MCP Server: Can be added to the configuration file of Claude Desktop or Cursor, or installed via Smithery (no local setup required).
- Antigravity (Google AI IDE): Added as a remote MCP server, with no authentication required, using serverUrl (not url).
- PyPI Package: deadends-dev.
- API: /api/v1/index.json.
- Website: deadends.dev.
- GitHub Repository:https://github.com/dbwls99706/deadends.dev。
Suitable Scenarios
- When an AI coding assistant encounters common code errors, query fixes that should not be tried.
- AI applications involving cross-border business, travel, law, or cultural compliance that need localized accurate answers.
- Developers building agent workflows that need to avoid known failures.
Limitations
- It does not cover all errors; missing cases can be quickly added via issue/PR/report_outcome.
- By design, it prioritizes practical solutions (dead end/workaround) rather than in-depth explanations.
- Reliability varies by domain/case; for high-risk changes, cross-validation with official documentation/vendor guidance is recommended.
- Country entries are manually reviewed (review_status: human_reviewed), with no LLM batch generation.
- Automatically collected data is candidate data, and is ultimately included after review by maintainers.
Version and License
- Current version v0.10.0, with country coverage expanded to 52 countries.
- Code is under the MIT License, and data is under the CC BY 4.0 License.
- Benchmark: 20 error scenarios, Precision@1=90%, MRR=0.935.
