Advanced code intelligence for AI coding agents.
Entroly builds bounded, source-verified repository understanding before context reaches the model: ASTs, Tree-sitter structure, typed graphs, interprocedural flow, architecture intelligence, semantic change analysis, verified refactors, and proof-carrying context.
Repository understanding beyond text similarity.
Similarity search can find text that looks related. Code intelligence has to preserve relationships the repository actually contains—and disclose when a relationship cannot be proven.
AST + parser intelligence
Native Python AST plus an optional open Tree-sitter registry for parser-backed declarations, recognized calls, exact byte spans, signatures, and structural profiles.
Typed repository graph
Symbols, imports, calls, containment, dependencies, reverse impact, neighborhoods, shortest paths, and relatedness with unresolved and ambiguous evidence kept visible.
Interprocedural reasoning
Source-verified Python argument→parameter and explicit return→result summaries for uniquely resolved static calls, with exact evidence spans.
Architecture intelligence
SCCs and cycle witnesses, condensation layers, communities, entry-to-foundation routes, hotspots, architecture diffs, and deterministic tie-breaks.
Semantic change intelligence
Git-object semantic diffs, portable graph snapshots, source hashes, cache commitments, freshness validation, and incremental invalidation reduce stale repository reasoning.
Build and test topology
Verified manifest ancestry across Cargo, Python, Node, Go, Zig, CMake, Maven/Gradle, Bazel, Swift, Elixir, Ruby, PHP, Dart, Clojure, Haskell, Meson, Make, and more—without pretending ancestry proves build inclusion.
LSP-enriched intelligence
An operator-configured language server can contribute bounded definitions, references, overrides, and workspace relationships; Entroly verifies source ranges before treating them as evidence.
Verified refactoring
Two-phase rename, safe-delete, and Python module-move plans bind edits to exact preimages, reject stale or tampered plans, preserve ambiguity, validate staged syntax, and expose rollback failures.
Evidence-aware context
Structural evidence is selected under an explicit token budget, with exact source identities, protected dependency signatures, recoverability, omissions, and tamper-evident receipts.
Local-first trust
Repository analysis stays local by default. Parser acquisition is explicit, air-gap mode wins, and external-process network behavior is labeled rather than silently claimed safe.
Evidence first, superlatives second.
Entroly is engineered to compete at the frontier of open-source code intelligence for AI agents, but the repository separates execution proof from source inspection and preserves negative results.
39-dimension conformance protocol: structural correctness, typed dispatch, caller/callee traversal, control/data flow, LSP UTF-16 ranges, repository hubs, rare-symbol ranking, cache invalidation, stale-source rejection, architecture reasoning, refactor safety, tamper evidence, and bounded graph queries are specified separately rather than collapsed into one marketing score.
Verified Code Context
Read the design and limitations for parser evidence, graphs, architecture, flow, LSP enrichment, receipts, and safe mutation contracts.
Conformance protocol
Inspect all 39 dimensions, their fixtures, execution statuses, and reporting rules.
Code intelligence FAQ
Is Entroly only a context compressor?
No. Compression is one layer. Entroly also builds local repository intelligence over AST/parser structure, typed call and dependency relationships, architecture, changes, flow, freshness, refactoring plans, provenance, and evidence budgets.
How is Entroly different from ordinary code search or RAG?
Search and RAG primarily rank textual similarity. Entroly can combine lexical relevance with verified structural relationships such as definitions, calls, imports, typed dispatch, dependencies, architecture, changes, and exact source freshness.
Does Entroly support multiple programming languages?
Yes. The parser layer is registry-open rather than hard-coded to a small language list. Python has the strongest verified semantic tier because Entroly also performs Python-specific AST/type and interprocedural analysis; other languages use available parser-backed structural evidence and conservative fallbacks.
Does Entroly guess when a call is ambiguous?
No by design. Ambiguous candidates and unresolved relationships remain explicit instead of being promoted into verified facts.
Can an AI agent use this intelligence under a context limit?
Yes. Repository relationships feed the same evidence-selection boundary that budgets context, preserves critical signatures and exact source spans, records omissions, and supports recovery.