The article promotes MCP server codebase intelligence from Pharaoh, giving AI agents architectural truth via a live code graph rather than blind scraping. It lists nine upgrades: knowledge-graph mapping, function search, blast-radius, reachability, dead-code detection, vision-gap analysis, consolida
Codebase Knowledge Graph MCPs convert a codebase into a precomputed, deterministic graph of facts (functions, imports, env vars) that AI agents query for architectural context. This reduces hallucinations, lowers token costs, and enables privacy-preserving local or tenant-isolated hosting for small
Context window limits degrade AI coding quality, causing missed files and duplicated logic as you reach practical tokens (4K–8K) and multiple sources. Bigger windows don’t guarantee better results; context rot and recency bias erode recall before the cap. The piece argues for architectural intel
The article argues that AI coding tools fail when they only treat code as isolated text; true reliability comes from giving agents full codebase context via knowledge graphs and a Model Context Protocol (MCP). It promotes building a code knowledge graph (Neo4j) from TS/Python to map functions, modul
AI coding tools struggle when they only read files, missing architecture, dependencies, and cross-service links in large repos. It promotes deterministic parsing and knowledge graphs to give LLMs true context and unlock graph-based queries. Pharaoh and MCP enable live, queryable repo graphs that int
AI coding assistants that see files in isolation breed architectural drift, duplicates, hidden blast radii, and unsafe suggestions. A structure-first workflow builds a live codebase map with module profiles, blast-radius and reachability graphs, and global function searches; enforces spec-to-code al
Claude code context windows often hit a limit, causing forgotten files, duplicated code, and dropped context during real work. Expanding the window isn’t a fix: code is token-heavy, histories accumulate, and latency and cost rise. The cure is disciplined context: modular sessions, /clear and /compac
AI tools can only see fragments of a codebase due to context limits and no system-wide view, causing duplicate utilities, broken callers, and hidden dependencies. Pharaoh turns TS/Python into a live Neo4j knowledge graph (Tree-sitter) and exposes it to AI tools via MCP, enabling blast-radius, reacha
The article argues that AI coding tools often rewrite or duplicate code because they only see partial context and misinterpret intent. Linearly reading files, missing architecture, and vague prompts cause regressions, dead code, and hidden dependencies. It advocates using knowledge graphs to map mod
AI tools often duplicate code because they only see local files, not the whole project, creating maintenance headaches for small teams. The article argues for architectural awareness—giving AI access to the entire codebase to reuse canonical logic. It highlights Pharaoh, a knowledge graph that maps
New posts on codebase intelligence, AI coding, and building Pharaoh.