Keryn Web
Scan a repository, a folder or a zip in the browser with no install, read the result, and export what you need. Keryn Web, at keryn.app, is the whole scanner as a static page. The scan runs in a worker inside the tab, Deep (the data-flow engine, compiled to WebAssembly) included. Local folders and files are read in the browser and never uploaded; a repository is read from its host, or through api.keryn.app without storage if you choose the fast archive.
Before you start
Section titled “Before you start”- A current Chrome, Edge or Firefox. Nothing to install and no account.
- Something to scan: a public repository link, a folder, a zip, or an example from the Demo menu.
- Optional: a GitHub token, which raises the request limit from 60 to 5,000 an hour (see Settings below).
What you can scan
Section titled “What you can scan”| Input | How |
|---|---|
| A public GitHub or Bitbucket Cloud repository | paste its link or owner/repo |
| Several repositories, or a whole organisation | paste a list, one per line, or github.com/org; see organization scans |
| A folder | choose it or drop it; it is read in place and read-only |
| A zip, or one file | drop it; Keryn detects whether it is source, a scanner report, a binary or a workflow export |
Workflow exports and binaries found inside a folder or zip are analysed with it, and their findings join the same result. Supported exports are n8n, Flowise, Langflow, Sim Studio, MCP, OpenAPI, CrewAI, LangChain and OpenAI Agents; binaries get the triage described under agent workflows and binaries.
Not sure where to start? The Demo menu loads synthetic inputs written for Keryn’s tests: a risky AI route, an AI inventory, an agent repository (code, MCP config, an n8n flow and a helper binary together), a Flowise agent flow, a binary, the AI stack of a page and a network AI surface.
The result
Section titled “The result”Scans starts an analysis and lists recent local evidence. Selected evidence opens Decision, with Chains, Findings, AI-BOM, Controls and Run details views. Organizations manages saved groups; Review queue keeps cross-repository decision work separate from the selected scan.
Choose Save in the result toolbar (Save to organization…) to pick an existing organization or name a new one. Opening or cancelling the dialog does not save. Evidence stays in this browser as-is, and no scan runs.
Use Scans → Artifact / Forge for binary capability inventory and Forge code-flow analysis. Standalone workflow/artifact evidence records its input identity and importer; repository checks are inapplicable, and its Decision view says Policy not evaluated.
The selected scan’s other views include:
-
Findings, sorted by severity, proof and file, and grouped by rule when there are more than 50. Each shows its proof level and the code around it. One workflow or binary issue is one finding, even when the analyser reports it several ways.
-
AI-BOM, key assets first: models, services and prompt files, then SDKs, each with the findings in the files that use it. Show all lists the rest. Components of known vendors carry a trust card built offline from the vendor’s public pages; opening it makes no request.
-
Agent surface (routes → tools): the edges from HTTP routes that reach a model to the tools they can call, with write, send, exec and payment tools highlighted. Routes are matched within one file.
-
Repository facts for a GitHub repository: licence, releases, contributor count and health. Loaded on request (four GitHub requests), or automatically if you pasted a token.
-
Investigate with AI: without a key, opens signed examples for all four Try one repositories, with pinned commits and review methods. These recorded examples send no request and change no verdict or gate. Provider recordings and Codex source reviews are labelled; upheld inventory observations confirm references, not vulnerabilities. A configured model offers a separately confirmed live second opinion; see connectors and agentic triage. The Trust Canvas draws an edge model of a scan.
-
See the trust graph (Trust graph on a phone): derives a trust graph from the scan (routes, model call sites, providers, tools with their authority, MCP servers, secrets by kind and place, data stores) and opens it at
#/trust, where the engine finds boundaries, chains and chokepoints. It is built in the tab from what was scanned, and nothing is uploaded. In an organization, the same button on the dashboard draws one zone per repository with shared providers and MCP servers. See start from a repository. -
Controls: the findings mapped to OWASP LLM and Agentic, MITRE ATLAS, NIST AI RMF, ISO/IEC 42001, the EU AI Act, NIST SP 800-53, ISO/IEC 27001 and CWE, each control marked findings, clear or not tested. References, not conformance claims: see controls, posture and hand-off reports.
Deep reports its progress, the files it took and the ones it left out, and Stop Deep keeps the pattern-rule results. If Deep cannot start or fails, the Results view says “Cross-file analysis did not run, so no flow between files is shown” and gives the reason; a result without cross-file proof is then at proof 1 and 2 only. A zip keeps every path it carries, so files at its root and several top-level folders resolve their imports.






Export
Section titled “Export”The Export menu generates files in the tab:
| File | Use |
|---|---|
| SARIF 2.1.0 | GitHub code scanning and other tools; see formats |
| Result JSON | the full report |
| Evidence bundle | findings with quoted evidence and hashes; see evidence bundles |
| Problems (VS Code) | path:line:col lines with a problemMatcher on top, for the Problems panel |
| CycloneDX 1.6 AI-BOM | the bill of materials |
| Hand-off reports | the Reports menu: control gap assessment, evidence pack, agent tool permission matrix and posture report, each one self-contained HTML (no script, nothing remote) that you print to PDF |
Developer tools
Section titled “Developer tools”Open Developer tools in the header. In-app tools contains API Workbench and Agent Inspector; external documentation links have their own group.
API Workbench
Section titled “API Workbench”At #/api, Load example imports a packaged synthetic OpenAPI document and immediately shows the reviewed suite JSON. The example describes a local API at http://127.0.0.1:8787; loading it starts no server and sends no requests.
Author a suite in three steps:
- Import routes: supply the application ID, build revision and explicit target origin. Optionally import OpenAPI 3.0/3.1 or HAR, then select a route.
- Identity and working control: describe the authenticated actor and a successful control request. JSON pointers are under Advanced.
- Security case: supply the request and expected response. Extra assertions and trial counts are under Advanced.
Each step validates before continuing. Validate reviewed suite checks structure and authored identity, control and security assertions; it does not prove target behavior. The suite JSON preview appears beside the form on wide screens and below it on phones. Download the reviewed suite for keryn api run, or explicitly run against an allowed target from the browser when that target allows the page’s origin through CORS. Opening the tool, loading the example and validating the suite send no target requests.
Agent Inspector
Section titled “Agent Inspector”At #/agents, Load example immediately inspects a packaged MCP configuration and SKILL.md, labelled Example. The result shows candidate files, declared access, review signals and evidence limits. It starts no MCP server, executes no skills and contacts no tools. Effective runtime authority stays unknown.
To inspect your project, supply a stable project ID and choose files or a folder. Advanced: compare or open saved inspections holds the optional earlier baseline and saved CLI/browser inspection inputs. Download the inspection JSON or review report. The CLI counterpart is keryn agents inspect; see audit and guard MCP servers. Browser file selections and CLI project discovery have separate comparison scopes.
Coming after 0.5.2
Section titled “Coming after 0.5.2”Settings
Section titled “Settings”Settings has three tabs.
- General: an optional GitHub token (60 requests an hour without one, 5,000 with), the folders never entered, the largest file read (5 MB by default, up to 25), the most files per repository (up to 200,000) and the theme. Settings are kept in your browser; the token stays in memory unless you choose to remember it for the tab.
- Import: read another scanner’s report, or analyse a workflow export, a binary, a page snapshot, observed ports or a runtime policy. See import.
- Connectors: the models and data stores Keryn may send findings to, only when you ask. See connectors.
Posture and Cloud
Section titled “Posture and Cloud”- Posture (
#/posture) is a questionnaire of 12 domains that the scan pre-fills with suggestions and caps. Your answers stay in your browser. See controls, posture and hand-off reports. - Cloud (
#/cloud) is the optional hosted part: sign in with your email and scan a website, a host’s AI services or a remote MCP server from Keryn’s servers. It receives the address you type, never your code or results. It is documented in its own section: Keryn Cloud.
What the page sends
Section titled “What the page sends”Until you sign in to Cloud, the only requests the page makes on its own behalf are the ones that fetch the repository you named, from api.github.com, raw.githubusercontent.com or api.bitbucket.org, without cookies. The privacy page lists every call, including the connector calls that start only when you press Test, Triage or Send.
What you just proved
Section titled “What you just proved”Open your browser’s developer tools, Network tab, and scan a folder: no request leaves the page. Scan a public repository: the requests go to the host that serves it and nowhere else. Cloud is the exception you opt into: its requests go to Keryn’s sign-in and api.keryn.app, and only from the Cloud page. The scan is an edge from your files to a worker in your own tab, and the page’s content security policy makes any other destination impossible until you press a connector button. The privacy page lists every call, and a test compares it with the code’s own list.
- Several repositories at once: organization scans.
- A second opinion on findings: triage.
- Keep the results: evidence bundles and export to Supabase and FalkorDB.
- The same scan in a pipeline: CI.
Progressive review and Scan details
Section titled “Progressive review and Scan details”Initial findings remain provisional while required work runs. Selection, focus, expanded evidence and row order are retained across updates; newly arriving findings append and are announced politely in batches. The release assessment remains pending until required work completes. Stop scan stays visible without opening details.
Scan details shows lane bars and a text table. A current trace records explicit predecessor edges and highlights a longest recorded core dependency path. Task durations include lane waits; nested worker spans, pre-core downloads and unmeasured gaps are excluded from that path weight. Partial/interrupted or invalid traces have no completed-path claim. Old interval-only reports retain a visibly labelled interval estimate. This measures scheduling, not security or CPU utilization.