`vulkro mcp serve`

Runs Vulkro as a Model Context Protocol server, so an AI coding agent can scan your code, read the proof behind a finding and check a fix before applying it.

vulkro mcp serve              # stdio (default)
vulkro mcp serve --port 9100  # SSE on http://127.0.0.1:9100/sse

Set up in Claude Code, Cursor and Codex

Claude Code: register the server once.

claude mcp add vulkro -- vulkro mcp serve

Cursor: add this to ~/.cursor/mcp.json, then restart Cursor.

{
  "mcpServers": {
    "vulkro": { "command": "vulkro", "args": ["mcp", "serve"] }
  }
}

Codex uses the Vulkro skill instead of the server. The installer appends instructions to ~/.codex/AGENTS.md, and Codex runs the vulkro CLI directly:

curl -fsSL https://dist.vulkro.com/skill-install.sh | bash -s -- --agent codex

The same installer takes --agent claude-code or --agent cursor (see Claude Code skill).

VS Code agent mode: the VS Code extension registers the server for you.

Other MCP hosts

Claude Desktop, Windsurf and Continue use the same mcpServers JSON as Cursor, in their own config file:

Host	Config file
Claude Desktop	`~/Library/Application Support/Claude/claude_desktop_config.json` (macOS), `~/.config/Claude/claude_desktop_config.json` (Linux), `%APPDATA%\Claude\claude_desktop_config.json` (Windows)
Windsurf	`~/.windsurf/mcp.json` (older versions: `~/.codeium/windsurf/mcp_config.json`)
Continue	`~/.continue/mcp.json`
VS Code MCP client	`~/.vscode/mcp_settings.json`

Restart the host after editing.

Tools exposed

Every tool is deterministic, calls no model and never writes to your repository.

Tool	What it does
`scan_project({path, format?})`	Full scan. Returns findings and a `scan_id`. Agents should pass `format: "compact"`.
`scan_file({path})`	Findings for one file.
`scan_diff({path, base_ref?})`	Findings on the lines a git diff added or changed (default base `HEAD~1`).
`get_findings({scan_id, severity?, finding_ids?})`	Re-filter a prior scan without rescanning.
`prove({finding_id or file+line, scan_id or path})`	The hop-by-hop proof for one finding. An empty chain means unproven, never safe.
`explain({rule_id})`	Markdown explanation of a rule.
`list_rules()`	The rule catalogue.
`suggest_fixes({scan_id, finding_id?, verify?})`	Fix patches, each re-scanned on a temporary copy: `fixed`, `not-fixed` or `regressed`.
`verify_fix({diff, scan_id or path, finding_id?})`	Apply your diff to a temporary copy, re-scan, return the same verdict.
`verify_sarif({sarif, path})`	Mark another scanner's SARIF results `proven`, `not-provable` or `not-examined`.
`inspect_repo({path})`	Review a just-cloned repo for malicious-code patterns. Never a "safe" verdict.
`read_code({path, target})`	Read `FILE`, `FILE:NAME` or `FILE:START-END` without reading a whole file.
`check_change({path, base?, fast?})`	After an edit: broken call sites, removed names still used, tests to run, findings on changed lines.
`assess_change({path, symbol?, base?})`	Blast radius before an edit, or dependents, tests and new findings after one.
`code_graph({path, mode?, ...})`	Scan-free code graph: file outline, usages of a name, one function's body, callers.
`from_trace({path, trace})`	Map a stack trace onto symbols and findings.
`evidence_graph({path})`	The `evidence-graph/1.0` document for a project.
`aggregate({graphs})`	Link evidence graphs across repositories.

Without an active trial or Pro, each tool call fails with JSON-RPC error -32001 and reason: "license_required". The server keeps running.

The server also exposes the local AI memory under .vulkro/ai/ as read-only resources (vulkro-ai-memory:///<id>).

AI-usable evidence-graph workflow

Call evidence_graph({path}) once per repository.
Pass the documents to aggregate({graphs}).
Read the returned cross-repo links, each with the callee's auth posture.

Vulkro supplies the evidence; the AI coding agent does the reasoning.

Flags

Flag	Description
`--port <PORT>`	Use SSE on `127.0.0.1:<PORT>` (`/sse` and `/messages`) instead of stdio.
`--offline`	Block every outbound network call.

Variable	Effect
`VULKRO_MCP_LOG=1`	Trace lines on stderr.
`VULKRO_OFFLINE=1`	Same as `--offline`. Other `VULKRO_*` scan variables also apply.

Exit codes

0 clean shutdown. 1 a refused tool call the caller treated as fatal. 2 bind failure, IO error or bad argument.

Set up in Claude Code, Cursor and Codex​

Other MCP hosts​

Tools exposed​

AI-usable evidence-graph workflow​

Flags​

Exit codes​