SAST and secrets scanning, with AI triage that shows its work.

OpenRouting runs Semgrep on your repository in a sandbox with no network. Then Claude, via Amazon Bedrock, reads each finding next to its code and tells you whether it's real, why, and what the fix looks like. Every verdict comes with its reasoning, so you can check it rather than trust it.

High · CWE-89

Raw SQL query executed

app/app.py:265

259search_term = content['search']
260print(search_term)
261str_query = "SELECT first_name, last_name, username FROM customer WHERE username = '%s';" % search_term
⋮
265search_query = db.engine.execute(str_query)

Real issue. The search value comes straight from the request body and is formatted into the SQL string that line 265 executes. Sending ' OR '1'='1 returns every customer.

261str_query = text("SELECT … WHERE username = :u")
265search_query = db.engine.execute(str_query, {"u": search_term})
Illustrative verdict on real code from the scan below.
scan started  github.com/we45/Vulnerable-Flask-App @ master
✓ clone      done  b6a4f97
✓ sandbox    network=none  fs=read-only  caps=dropped
✓ semgrep    29 results
✓ dedupe     28 findings
HIGH CWE-89   app/app.py:265  sqlalchemy-execute-raw-query
HIGH CWE-502  app/app.py:329  insecure-deserialization
HIGH CWE-96   app/app.py:103  dangerous-template-string
HIGH CWE-250  Dockerfile:15  missing-user
+ 24 more findings
done in 18.4s  8 high  20 medium  clone deleted
> 

Real results from scanning we45/Vulnerable-Flask-App, an intentionally vulnerable open-source app, laid out as a terminal. Scans start from the dashboard, a pull request or your AI assistant.

Where you already work

Review the pull request before it merges, triage in the dashboard, or ask Claude or ChatGPT about your findings over MCP.

Pull request security review

Pick an open pull request and OpenRouting reports findings only in the files it changes. Its findings stay on that review and never touch the repository's tracked list. When you're ready, one summary comment goes to the PR — posted as you, when you click, never automatically. Pull request review →

  • ScopeFindings in the files the PR changesdiff
  • TriageLikely false positives left out of the commentAI
  • CommentOne summary, as you, on clickGitHub
  • CostPer review, not per seat10 credits

A dashboard built for triage

Filter by severity, scanner, status or file, move with J and K, and set a status with one key. Bulk-select, undo, and share a link to exactly the view you're on. A trend chart shows open findings by severity over time.

Ask Claude or ChatGPT about your findings (MCP server)

Connect OpenRouting as an MCP server and your assistant can search findings, read a verdict and its fix, set a finding's status or start a scan — after you sign in and approve the access it asks for. Reads are free; scans cost the usual credits. Text taken from scanned code is labelled untrusted, and known key and token formats and secret-named values are redacted before it reaches the assistant.

Claude Code: claude mcp add --transport http openrouting https://openrouting.si/mcp. Claude: add https://openrouting.si/mcp as a custom connector. ChatGPT: Developer mode, on plans that offer it. MCP setup and tools →

$ claude mcp add --transport http openrouting https://openrouting.si/mcp
> Which high-severity findings are open in Vulnerable-Flask-App?
⏺ openrouting · search_findings  severity=[high] status=[open]
8 open · CWE-89 app/app.py:265 · CWE-502 app/app.py:329 · +6 more
> Show me the SQL injection one.
⏺ openrouting · get_finding  verdict · explanation · suggested fix
Real issue: the request's `search` value is formatted into the SQL string.
Fix: bind it as a parameter — text("… WHERE username = :u").
> Fixed on master. Re-scan it.
⏺ openrouting · start_repository_scan  20 credits — approve?
Illustrative exchange. The install line and tool names are real; the findings are the ones from the scan above.

What it adds on top of the scanners

Open-source scanners find the candidates. These are the parts that make the result something you can act on, and the mechanism behind each.

  • Your AI-agent setup is attack surface. It gets reviewed.

    When a repository has CLAUDE.md, AGENTS.md, .claude agents, skills, hooks or settings, or MCP configs, a full scan inventories them and Claude reviews them for prompt-injection paths, over-broad tool permissions and exposed secrets. The setup generator can then open a pull request with a hardened setup — tight tool allowlists, deny rules, out-of-repo imports removed — for you to review.

    +10 credits, only when a setup is found · setup PR 50, never auto-merged

    AI-agent security →
  • Known secret formats are redacted before any model sees them.

    Semgrep's secret rules flag cloud keys, tokens and passwords in the files you scan. Known key and token formats and secret-named values become [REDACTED] before any AI model or assistant sees them; the finding in your dashboard still shows the flagged line so you can fix it. The clone is deleted when the scan ends.

    one redactor before triage, reports, the agent-setup review and MCP results

    How code is handled →
  • API collections checked without touching your API.

    Upload a Postman, OpenAPI 3, Swagger 2, HAR, Insomnia or Bruno file. Each endpoint is checked for missing or weak auth, secrets and PII in URLs, guessable IDs, mass assignment and responses that over-share — from the document alone. Header values and raw example values are discarded on import.

    no request sent to your API · 20 credits per analysis

    API suite analysis →
  • Website scans you can point at production.

    OWASP ZAP runs passive checks only: no attack payloads, no form posts. Before every scan the address is resolved and refused if it is private, loopback or a cloud-metadata IP; the scanner is pinned to that one IP, and the crawl is held to the host with off-host redirects stopped.

    passive only · 10 credits · active scanning is not offered yet

    How passive scans work →
  • Each issue tracked once, across every scan.

    Findings are fingerprinted, so a re-scan updates the existing finding instead of adding a duplicate. When a complete scan no longer sees it, it closes as fixed; if it comes back, it reopens. Findings you mark as false positive or accepted risk stay that way unless the issue gets worse, with who changed them and why.

    states: open · false_positive · accepted_risk · fixed

    How it works →
  • A report you can hand to someone else.

    For a repository or website scan, Claude writes the narrative and OpenRouting adds the tables: findings by severity, scanner and CWE, the most-affected files or pages, every finding, and detail for each high and critical.

    Markdown or PDF · 10 credits, refunded if it fails

    All features →

What it checks today

Only what ships. Code, secret and IaC rules come from Semgrep's default ruleset; websites are checked by OWASP ZAP in passive mode. Not covered yet: dependency CVEs (SCA), active website testing and secrets in past commits.

  • Code flaws (SAST)

    Semgrep

    Semgrep's default ruleset: SQL and command injection, XSS, unsafe deserialization, path traversal and more, in Python, JavaScript/TypeScript, Java, Go, Ruby, PHP, C# and others.

  • Hardcoded secrets

    Semgrep

    Semgrep's secret rules flag cloud keys, tokens and passwords in the files you scan. Values are replaced with [REDACTED] before any AI call.

  • Infrastructure as code

    Semgrep

    Semgrep's rules for Terraform, Dockerfiles, docker-compose, Kubernetes manifests and GitHub Actions workflows.

  • Live websites (passive DAST)

    OWASP ZAP

    OWASP ZAP baseline in passive mode: security headers, cookie flags, information leaks and mixed content. No attack payloads, no form posts.

  • API collections

    API rules

    Postman, OpenAPI, Swagger, HAR, Insomnia and Bruno imports checked for missing auth, secrets or PII in URLs, guessable IDs, mass assignment and over-shared responses. No traffic is sent.

  • AI-agent setups

    Claude review

    CLAUDE.md, AGENTS.md, .claude agents, skills, hooks and settings, MCP configs and the agent frameworks named in your dependencies, reviewed for prompt injection, unsafe tools and exposed secrets.

21 specialist review agents

Each finding is routed by scanner, CWE and rule to one of these. Its review guidance shapes the verdict and fix, and the report names which agent reviewed it.

What each agent checks →

How OpenRouting handles your code

A security tool reads your most sensitive asset. This is what happens to it. Full security details →

  • Your code is never run

    No dependency installs, builds or scripts. Scanners read source files and nothing else.

  • Code scanners have no network

    Each code scan runs in a fresh container: networking off, repository mounted read-only, non-root, all capabilities dropped, CPU, memory and time capped.

  • Clones can't run repository config

    Git hooks, filters and credential helpers are neutralised, submodules are not followed, symlinks are flattened and clone size is capped.

  • Secrets redacted before AI

    Claude, via Amazon Bedrock, receives the lines around each finding; your AI-agent config files, such as CLAUDE.md and .claude settings, when the agent-setup review runs; the files the setup generator opens (400 KB at most) when you run it; a findings digest when you generate a report; and, for website and API-collection scans, each finding's scan evidence or endpoint details. Never the whole repository. Known key and token formats and secret-named values are redacted before any AI model or assistant sees them.

  • Clones deleted after every scan

    OpenRouting keeps the flagged lines it shows you, not a copy of your repository.

  • Tokens encrypted, workspaces isolated

    GitHub tokens are encrypted at rest and never sent to the browser. Every request is checked against your workspace.

Pay per scan, not per seat

100 free credits to start · full repository scan 20 · pull request review 10 · website scan 10 · $1 per 20 credits · failed runs refunded. Every price →

Common questions

What does OpenRouting scan?

Repository code for flaws (SAST), hardcoded secrets and infrastructure-as-code, using Semgrep's default ruleset; pull requests, reporting findings only in the files they change; live websites, passively, with OWASP ZAP; API collections from Postman, OpenAPI, Swagger, HAR, Insomnia or Bruno files; and AI-agent setups such as CLAUDE.md, .claude settings and MCP configs. It does not scan dependencies for known CVEs (SCA) yet.

Which languages are supported?

Code scanning uses Semgrep's rules, which cover Python, JavaScript and TypeScript, Java, Go, Ruby, PHP, C#, Kotlin and more, plus Terraform, Dockerfiles, Kubernetes manifests and GitHub Actions workflows.

How does AI triage decide what is a false positive?

Claude, via Amazon Bedrock, reads each finding with the code around it and the guidance of the specialist agent for its CWE or rule — SQL injection, SSRF, access control and 18 others. It returns a verdict (real issue or likely false positive), a confidence level, a plain-English explanation and a suggested fix. Nothing is dismissed for you: you see the reasoning and set the status.

Is my code sent to an AI model?

Claude, via Amazon Bedrock, receives the lines around each finding; your AI-agent config files, such as CLAUDE.md and .claude settings, when the agent-setup review runs; the files the setup generator opens (400 KB at most) when you run it; a findings digest when you generate a report; and, for website and API-collection scans, each finding's scan evidence or endpoint details. Never the whole repository. Known key and token formats and secret-named values are redacted before any AI model or assistant sees them. The clone is deleted when the scan ends; OpenRouting keeps each finding's flagged lines and nothing else from your code.

Can I use it from Claude, ChatGPT or Claude Code?

Yes, through the MCP server at https://openrouting.si/mcp. Add it as a custom connector in Claude (Settings → Connectors), in ChatGPT with Developer mode on plans that offer it, or in Claude Code with: claude mcp add --transport http openrouting https://openrouting.si/mcp. Claude and ChatGPT use sign-in: you approve read and optionally write access and can revoke it in AI assistants (MCP) → Connected apps. Other tools, such as Cursor, use your personal API key from Dashboard → AI assistants (MCP), sent as an Authorization header; it can be made read-only, disabled or regenerated there. Reading findings is free; starting a scan costs the usual credits.

Can OpenRouting scan a running website?

Yes, passively. OWASP ZAP checks security headers, cookie flags, information leaks and mixed content using ordinary page loads, with no form submissions or attack payloads. Active scanning is not offered; it will require proving you own the domain.

What happens to findings I mark as false positives?

They leave your open list and stay out on later scans unless the issue gets worse, with a record of who marked them and why. You can reopen any finding at any time.

Do I need to install anything?

No CLI or CI step. Sign in with Google or GitHub, connect a repository and start a scan from the browser. The MCP connection is optional.

See which of your findings are real.

Connect a GitHub repository and get triaged results in minutes. A full scan costs 20 of your 100 free credits.

Start with 100 free credits