Maps architecture, risk, and runtime surfaces — from 73-file libraries to 25,432-file infrastructure repos in under a minute. Entirely offline, zero dependencies, 100% open source. AI-ready context packs, not vague summaries.
Seven structured outputs from a single scan — no cloud, no uploads, no dependencies beyond Python stdlib.
Live directed graph of your codebase — entry points, module boundaries, dependency flows, resolved through a 5-tier ranked fallback.
Maintainability, runtime complexity, test signal, security — with a detailed breakdown so you know exactly where the pain is.
Identifies files with highest risk — oversize files, TODO density, doc drift, test gaps. Ranked by severity so you attack the worst first.
Compact ~2,500 token brief that replaces hours of file reading — ready for Cline, Claude Code, Codex, Roo, Continue. Copy, paste, ship.
Measured across four real-world repositories spanning three orders of magnitude — from 73 to ~40,000 files. Honest numbers, no cherry-picking.
Sentinel doesn't summarize. It builds a structured model — identity resolution, architecture graph, runtime awareness, risk scoring, then compresses everything into an AI-ready context pack (~2,500 tokens).
5-tier ranked fallback — known repo names, manifests, README body, headings — never returns "Sponsors" for FastAPI or garbage for Kubernetes.
Parses imports, modules, and dependencies into a directed graph. Classifies every file by role — runtime, test, build, docs, vendor, generated.
Computes maintainability, runtime complexity, test signal, and security signals per file. Flags hotspots: oversize files, TODO density, doc drift, test gaps.
Produces a ~2,500 token context pack with ranked next actions — ready for Cline, Claude Code, Codex, Roo, Continue. Copy, paste, understand.
Transparent about where Sentinel is still evolving. These are honest edges — not hidden flaws.
5-tier fallback resolves most cases (FastAPI → "FastAPI", not "Sponsors") but unconventional manifests can still confuse. Actively tuning tier weights.
Large monorepos with unconventional layouts may rank secondary entry points below expected thresholds. Go binaries get a +80 bonus — still tuning for edge cases.
Sparse or inconsistent docstrings can reduce docs-to-code alignment accuracy. Purpose inference uses 6-step fallback — works on well-documented repos, drifts on sparse ones.
Apache 2.0 licensed. Zero dependencies. No telemetry, no accounts, no cloud. Built for developers who want AI to understand their codebase — without uploading to the cloud.
Scan your repository in seconds. See what your codebase is actually telling you.