TRACE
Sign in
open source · self-hosted · MIT

Find real vulnerabilities.

Fix them only with your approval.

TRACE runs real security scanners on your repository, lets AI agents investigate each finding with cited evidence, and verifies every fix by re-scanning before it opens a pull request.

Watch the 90-second demo
  • T Threat
  • R Risk
  • A Analysis
  • C Compliance
  • E Evaluation
localhost:3000/scans · juice-shop @ 717f5ef
✓trivy89 findings
✓semgrep139 findings
✓gitleaks69 findings
✓osv-scanner0 findings*
288 unique20 critical101 high40 medium
SQL injectionHIGH

routes/login.ts:34

AI verdict: user input reaches a Sequelize query through string interpolation. Use parameterized queries.

risk 7.0 · P26 cited sourcesread via MCP
4/4 scanners ✓
verified by re-scan

A real TRACE scan of OWASP Juice Shop, summarized.* Juice Shop ships without a lockfile, so there were no dependency versions to check. TRACE does not flag this yet.

297 → 288

raw findings from that scan, deduplicated into one list

4

open-source scanners, side by side

4

intel sources: NVD, OSV, GitHub Advisories, CISA KEV

0

automatic merges. Every fix waits for you.

Trivy ✦OSV-Scanner ✦Semgrep ✦Gitleaks ✦NVD ✦OSV ✦GitHub Advisories ✦CISA KEV ✦LangGraph agents ✦MCP tools ✦Ollama ✦re-scan verified ✦human approval ✦audit trail ✦Trivy ✦OSV-Scanner ✦Semgrep ✦Gitleaks ✦NVD ✦OSV ✦GitHub Advisories ✦CISA KEV ✦LangGraph agents ✦MCP tools ✦Ollama ✦re-scan verified ✦human approval ✦audit trail ✦

Demo

See the whole flow in 90 seconds.

The product screens in the video come from a real run.

trace-psi.me/#demo

A 90-second narrated tour: four scanners merged into one deduplicated list, live vulnerability intelligence, a real AI investigation with cited evidence, a deterministic risk score, and a human-approved fix verified by re-scanning.

How it works

From raw alert to verified fix.

Five steps, each recorded in the audit trail.

  1. 01

    Connect

    A GitHub repo, any public git URL, or a local fixture. Pick a branch.

  2. 02

    Scan

    Trivy, OSV-Scanner, Semgrep and Gitleaks on an isolated snapshot. Repository code is never executed.

  3. 03

    Enrich

    Findings are deduplicated and matched against NVD, OSV, GitHub advisories and CISA KEV.

  4. 04

    Investigate

    LangGraph agents read the code through MCP tools and explain each finding with cited evidence.

  5. 05

    Approve & verify

    You approve or reject the fix. TRACE re-scans a patched snapshot, then opens a PR. It never merges.

Coverage

What TRACE finds.

Code vulnerabilities

SQL injection, command injection, unsafe deserialization and more, across about 30 languages including C and C++.

routes/login.ts:34 · SQL injection

Vulnerable dependencies

Known CVEs in Python, JavaScript, Java, Go, Ruby, Rust, PHP and .NET packages, with the fixed version.

pyyaml 5.3 → 5.4 · CVE-2020-14343

Leaked secrets

API keys, tokens and private keys committed to the repo. Values are redacted and never stored.

63 secrets in Juice Shop

Infrastructure

Terraform, Dockerfiles and CI workflows: open security groups, plain HTTP, shell injection in GitHub Actions.

networking.tf:86 · open egress

AI and agent apps

Prompt-injection and tool-misuse checks, plus live probes judged with a random canary.

leakage · forbidden tool calls

Honest by design.

A scanner that cannot run is reported as unavailable or skipped, never shown as a clean result.

Why TRACE

Scanners alone vs TRACE.

TRACE uses the same open-source scanners you already trust. It adds what happens after the alert.

Scanners alone

  • What you getHundreds of raw alerts, duplicated across tools
  • ContextA rule id and a line number
  • PriorityEach tool’s own severity
  • The fixYours to write and test
  • Blind spotsA skipped scanner looks the same as a clean result

With TRACE

  • What you getOne deduplicated list, ranked by risk
  • ContextCVE intel plus an AI investigation that cites its sources
  • PriorityA deterministic risk score from CVSS, KEV and reachability
  • The fixA proposed patch, your approval, verified by re-scan, opened as a PR
  • Blind spotsCode no scanner read never gets a passing score

FAQ

Questions, answered plainly.

Does TRACE change my code?

Not without you. Every fix needs a human approval, is verified on an isolated copy of the repo, and is opened as a pull request. Nothing is ever merged automatically.

Does my code leave my machine?

TRACE is self-hosted. With Ollama, scans and AI investigations run entirely on your hardware. Repository code is never sent to a hosted model unless you switch that on.

Which AI models does it use?

Any local model through Ollama, any OpenAI-compatible endpoint (vLLM, OpenAI and others), or Anthropic. You choose per organization in Settings.

Is it free?

Yes. TRACE is open source under the MIT license.

What can’t it do yet?

It does not find deep logic bugs, such as protocol state errors, that no scanner rule describes. Dependency scanning needs a lockfile, and TRACE does not yet warn you when one is missing. Small local models are slow on a CPU and tend to give guidance rather than a patch, and an AI investigation can still agree with a false positive, so review its verdict. TRACE is alpha software.

Quick start

Run it on your machine.

Needs Node 20+, git and Ollama. Docker Compose works too.

bash
$ git clone https://github.com/atique-ahmad-01/trace.git && cd trace
$ scripts/dev-setup.sh   # scanners, Python venv, npm install, .env
$ scripts/dev.sh         # API :8000 + console :3000
Star on GitHubRead the setup guideTry the hosted demo
TRACE

Find real vulnerabilities.

Fix them only with your approval.

Try the hosted demo