End-to-end flow analysis
Track user-controlled and sensitive values across functions, model boundaries, and high-impact sinks such as shells, SQL, URLs, files, and external destinations.
Codelytra follows untrusted input and sensitive data through models, agents, tools, retrieval, and MCP—then shows the exact path to privileged actions.
$ npx codelytra scan .
Analyzing 48 files · 17 AI security rules
Model output reaches command execution
1 critical path found842 ms
Codelytra models AI-specific trust boundaries without importing, installing, building, or executing the repository it scans.
Track user-controlled and sensitive values across functions, model boundaries, and high-impact sinks such as shells, SQL, URLs, files, and external destinations.
See source locations, propagation steps, confidence, severity, and focused remediation.
Recognizes common OpenAI, Anthropic, Vercel AI, LangChain, Google GenAI, and MCP patterns.
Export SARIF, JSON, Markdown, HTML, and stable fingerprints for baseline-aware policies.
Bring any OpenAI-compatible provider. AI enriches deterministic evidence—it never decides findings.
Find AI providers, tools, inputs, sensitive sources, and privileged sinks.
Propagate trust tags through functions, prompts, retrieval, and model output.
Report the exact path, safeguards observed, and actionable remediation.
Apply severity and confidence policies in CI with stable exit codes.
Node.js 22+ is the only runtime requirement. Static scans require no account, API key, network access, or telemetry.
npm install --save-dev @ezducate/codelytra
npx codelytra scan .
Optional smart features can be configured later with
codelytra provider setup.
Review tool permissions and model-controlled arguments before an agent gains production access.
Block newly introduced high-confidence paths and upload SARIF to code scanning.
Generate an AI Security Manifest and attack graph for architecture and risk reviews.
JavaScript and TypeScript applications that use LLMs, AI agents, RAG, tool calling, or Model Context Protocol servers.
No. Static analysis is local-first, makes no network requests, collects no telemetry, and never executes the target repository.
It identifies code paths where untrusted or retrieved content can influence models and privileged actions, including missing validation, authorization, or approval controls.
Yes. Codelytra provides stable policy exit codes, SARIF output, baselines, JSON reports, and configurable severity and confidence gates.
Codelytra is maintained by EZducate, an AI-powered platform building accessible tools for educators, families, and clinicians. We are sharing the security foundations we want the broader AI ecosystem to have.
Explore EZducateOpen source under Apache-2.0. Built for teams shipping AI responsibly.