Corpus
the immune system for your code
Try it now -- scan any GitHub repository
Paste a repo URL and Corpus builds the structural graph, runs security scanners, and shows findings.
Features
UNDERSTAND
Auto-scans your codebase and builds a structural graph. Every file, function, and dependency mapped in seconds.
WATCH
Hooks into Claude Code and Cursor via MCP. Intercepts broken changes and auto-fixes them before they land.
SHOW
A visual graph of your entire project. Green means healthy, red means broken. Know your codebase at a glance.
15 Defense Layers
Two tiers of protection. Code scanners catch file-level issues. Agent guardrails enforce behavioral safety.
Graph Contracts
Diffs against structural graph. Catches removed functions, deleted guards, broken exports.
CVE Patterns
30 vulnerability patterns mapped to real CVE IDs. SQL injection, SSRF, prototype pollution.
Secret Detection
API keys, tokens, private keys, database URLs, webhook secrets in source code.
Code Safety
eval(), exec(), innerHTML, disabled SSL, SQL concatenation, wildcard CORS.
Dependency Check
Hallucinated npm packages, typosquats, non-existent imports. Checks against 12K+ known packages.
Pattern Intelligence
Learned from 280 repos. Context-aware: eval() in webpack = suppress, eval() in route = critical.
Trust Scoring
Per-file and codebase-wide trust scores. 0-100 based on finding density and severity.
Injection Firewall
Scans external content for prompt injection before it enters LLM context.
Exfiltration Guard
Detects PII in outbound payloads — emails, SSNs, credit cards — and redacts.
Context Poisoning
Scans stored memory for poisoning signatures like "ignore previous instructions."
Cross-User Firewall
Enforces user context isolation. Prevents data from user A bleeding into user B.
Session Hijack
Detects rapid-fire events and timing anomalies that indicate automated injection.
Confidence Calibrator
Audits action logs for overconfidence and underconfidence per intent category.
Scope Enforcer
Checks if proposed action falls within developer-declared scope boundaries.
Undo Integrity
Classifies actions as reversible, best-effort, or irreversible before execution.
Real-time immune intelligence. Learned from 280+ open-source repos.
Previously Scanned Repos
Benchmarked against popular open-source projects
| Repository | Files | Nodes | Findings | Scan Time |
|---|---|---|---|---|
| honojs/hono | 362 | 1,567 | 69 | 107ms |
| drizzle-team/drizzle-orm | 966 | 4,874 | 37 | 334ms |
| trpc/trpc | 909 | 2,936 | 8 | 255ms |
| shadcn-ui/ui | 3,383 | 12,840 | -- | 933ms |
| calcom/cal.com | 7,508 | 22,794 | -- | 2.1s |
| prisma/prisma | 2,813 | 6,782 | -- | 642ms |
| t3-oss/create-t3-app | 178 | 322 | 0 | 73ms |
No more AI slop
Your AI writes code
Claude Code, Cursor, or any MCP-compatible tool generates changes to your codebase.
Corpus intercepts and evaluates
Deterministic policy evaluation powered by Jac catches regressions, broken imports, type errors, and structural damage.
Auto-fix or alert
Broken code is healed automatically. If it cannot be fixed, you see it instantly in the visual graph -- red nodes, clear diagnostics.
Corpus watches so you don't have to.
npm install -g corpus-cli10 Jac Walkers guarding your AI agent
Deterministic policy evaluation powered by Jac. No LLM opinions. No probabilistic guessing. Pure graph traversal that returns PASS, CONFIRM, or BLOCK.
Why Jac for Policy Evaluation?
Deterministic
LLMs are probabilistic -- ask the same question twice, get different answers. Safety policies must be deterministic. Jac walkers traverse a graph and return the same verdict every time.
Graph-Native
Jac is built around graphs. Policy evaluation is graph traversal -- walkers visit nodes, check conditions, and report verdicts. No ORM, no SQL.
Composable
Each walker is independent. Stack 10 built-in policies, then add your own custom walkers. Each one checks a specific concern -- no tangled if-else chains.
Action Safety
Blocks destructive actions universally
Scope Guard
Enforces action scope boundaries
Rate Guard
Rate limiting for AI actions
Confidence Calibrator
Detects AI overconfidence & underconfidence
Injection Firewall
Blocks prompt injection attacks
Exfiltration Guard
Prevents PII data exfiltration
Session Hijack
Detects automated session injection
Cross-User Firewall
Prevents cross-user data access
Context Poisoning
Detects poisoned memory chunks
Undo Integrity
Validates undo capability before execution