deadends.dev

What is deadends.dev?

deadends.dev is a deterministic, country-scoped knowledge base that stops AI agents from repeating known failures in code and the real world. It covers 2,089 code error entries across 51 domains and 250+ country-scoped dead ends across 52 countries, including visa rules, legal red lines, cultural taboos, and emergency numbers.

How to use deadends.dev?

Install via pip (pip install deadends-dev) and run deadends "error message" from the command line. Alternatively, integrate as an MCP server for Claude Desktop, Cursor, or Antigravity (remote HTTP) via configuration files or Smithery. Use the Python SDK with lookup, batch_lookup, and search functions for programmatic access.

Key features of deadends.dev

• 90% Precision@1 and 0.935 MRR benchmark
• 2,204 canon entries across 54 domains and 39+ countries
• 11 MCP tools: lookup_error, get_error_detail, batch_lookup, etc.
• Deterministic, sub-millisecond local regex matching (no API roundtrip)
• Community-validated success rates that feed back into data
• Primary-sourced country canons citing government sites and embassies

Use cases of deadends.dev

• AI agent avoids known-failed code fixes (e.g., CUDA OOM, ModuleNotFoundError)
• Agent skips incorrect country-specific advice (e.g., tipping in Japan, public affection in UAE)
• Agent handles error chains (e.g., error A leads to error B 78% of the time)
• Agent checks visa rules, legal red lines, and cultural taboos by country
• Agent queries emergency numbers, driving norms, and food safety per jurisdiction

FAQ from deadends.dev

What makes deadends.dev different from asking an LLM directly?

It is deterministic (same query, same answer, no hallucination), country-scoped (ID format includes country code), primary-sourced, community-validated, and sub-millisecond local matching.

How do I install and run deadends.dev?

Run pip install deadends-dev then deadends "error message" on the command line. For MCP integration, add the server config to ~/.claude/claude_desktop_config.json or install via Smithery.

Are there any limitations?

Not all errors are covered; missing cases can be added via issues/PRs or report_outcome. The system prioritizes practical workarounds over deep explanations. High-risk changes should be cross-verified with official documentation.

How do I resolve an "Unauthorized" error when using the MCP server?

Ensure you are using only one mode (local or remote), check that cwd in the config points to the actual path, test the local server directly, and if using Smithery, reinstall the server. Restart Claude Desktop afterwards.

What API endpoints are available?

Endpoints include /api/v1/match.json, /api/v1/index.json, per-country JSON, stats, NDJSON streaming, and OpenAPI 3.1 spec. All are accessible via the public API at deadends.dev.