Hookwarden — Webhook Integrity
io.github.Hookwarden/mcp
Documentation
The first MCP server doing webhook signature verification.
Local. Deterministic. Zero network. JS / TS / Python — 21 providers — 3-state verdicts.
What it does
@hookwarden/mcp exposes hookwarden's deterministic webhook-verification engine as MCP tools. AI coding agents — Claude Desktop, Cursor, Continue, and apps built on the Anthropic Agent SDK — get two tools:
scan_handler— scan pasted webhook handler code and get back a structured finding (verified/not-verified/manual-review) for Stripe, GitHub, Shopify, Twilio, and 17 other providers.verify_audit_chain— verify a hookwarden evidence pack's hash chain + KMS-signed Merkle roots offline (no network, no auth), returning{ valid, broken_at_row, signatures_verified }.
Runs entirely client-side. Zero network egress from the MCP process. Zero auth. The same engine the hookwarden CLI ships — same rule pack, same content hashes, end-to-end provenance.
Install in one command
npx @hookwarden/mcp init
That command detects your installed MCP clients (Claude Desktop / Cursor / Continue), prints a per-OS summary table, and writes the canonical config to each one. .bak files land alongside every config it touches. Restart the relevant client and scan_handler + verify_audit_chain appear in the tool list.
Pass --all for non-interactive use, --dry-run to preview, --force to overwrite an existing mcpServers.hookwarden entry, or --clients claude-desktop,cursor to limit the set.
Manual configuration
If you'd rather not run the init helper, write the config by hand.
Claude Desktop
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%/Claude/claude_desktop_config.json
{
"mcpServers": {
"hookwarden": {
"command": "npx",
"args": ["-y", "@hookwarden/mcp"]
}
}
}
Cursor
~/.cursor/mcp.json — same shape as Claude Desktop:
{
"mcpServers": {
"hookwarden": {
"command": "npx",
"args": ["-y", "@hookwarden/mcp"]
}
}
}
Continue.dev
Continue uses a per-server file. Write ~/.continue/mcpServers/hookwarden.yaml:
name: hookwarden
command: npx
args: [-y, "@hookwarden/mcp"]
Anthropic Agent SDK
The Agent SDK is programmatic — there's no config file to edit. Drop this into your code:
import { query } from "@anthropic-ai/claude-agent-sdk";
for await (const message of query({
prompt: "Audit this webhook handler for missing signature verification...",
options: {
mcpServers: {
hookwarden: {
command: "npx",
args: ["-y", "@hookwarden/mcp"],
},
},
},
})) {
console.log(message);
}
The init helper does not write here — the SDK integration is code, not config.
The scan_handler tool
Inputs
| Field | Type | Notes |
|---|---|---|
code | string | Single-blob mode. Defaults to TypeScript; override with language. Mutually exclusive with files. |
files | Record<string, string> | Multi-file mode. Keys are virtual file paths; language inferred from extension. |
language | "js" | "ts" | "python" | Optional. "php" returns language_not_in_preview — PHP support ships in v0.8.1. |
provider | string | Optional. Case-insensitive ("stripe" = "Stripe"). unknown_provider if not in the rule pack. |
Output
Returns Tool.result with structuredContent first-class (MCP spec 2025-06-18) plus a companion stringified-JSON text block for backwards compatibility:
{
"verdict_summary": { "verified": 0, "not_verified": 1, "manual_review": 0, "parse_error": 0 },
"findings": [
{
"rule_id": "stripe/missing-signature-verification",
"provider": "stripe",
"severity": "critical",
"verdict": "not-verified",
"file": "__handler.ts",
"line_start": 7,
"line_end": 9,
"message": "Webhook handler reads req.body without calling stripe.webhooks.constructEvent...",
"provider_docs_url": "https://stripe.com/docs/webhooks",
"references": [
"https://www.svix.com/blog/common-failure-modes-for-webhook-signatures/",
"https://hookdeck.com/webhooks/guides/webhook-security-vulnerabilities-guide"
],
"rule_pack_version": "0.7.2"
}
],
"scan_metadata": {
"engine_version": "0.7.2",
"rule_pack_version": "0.7.2",
"rule_pack_content_hash": "938f7565e29f19e0cbd34594731d1de2b0ffd40ee87aacaa8be2435fc1f47f32"
}
}
Verdict vocabulary
| Verdict | Meaning |
|---|---|
verified | Engine proved the handler verifies signatures via the provider's documented SDK call. |
not-verified | Engine proved the handler does NOT verify signatures (or verifies after side effects). |
manual-review | Engine has structural evidence but can't decide without human eyes — review the snippet. |
Same 3-state vocabulary as the CLI. No remapping in the MCP layer.
Error payloads (isError: true)
| Error | Trigger |
|---|---|
empty_input | Neither code nor files provided. |
exclusive_input_modes | Both code and files provided — pick one. |
language_not_in_preview | language: "php" — PHP ships in v0.8.1. Suggestion: npx hookwarden scan. |
unsupported_language | language outside the known set. |
unknown_provider | provider not in the rule pack. Response includes the sorted known list. |
engine_drift / rules_drift | Boot or call-time drift — see drift-detection. |
Parse failures don't crash the transport — they emit as rule_id: "engine/parse-error" findings with severity high and increment verdict_summary.parse_error.
The verify_audit_chain tool
Verify a hookwarden evidence pack's integrity offline — no network, no AWS, no credentials. "Verify the chain without trusting hookwarden." The tool walks the per-row hash chain, roundtrips the manifest hash, and — for v1.1 packs that embed the signing CMK's SPKI public key — verifies the ECDSA-P256 (MessageType:DIGEST) Merkle-root + manifest signatures with node:crypto.
Inputs
| Field | Type | Notes |
|---|---|---|
evidence_pack_json | string | The full evidence-pack JSON (the file hookwarden's evidence export produces). Both v1.0 and v1.1 packs are accepted. |
Output
{
"valid": true,
"broken_at_row": null,
"signatures_verified": 2
}
| Field | Meaning |
|---|---|
valid | true iff the hash chain is intact AND the manifest hash roundtrips AND every signature that could be verified offline verified. |
broken_at_row | The seq of the first row whose chain hash didn't match, or null if the chain is intact. Signature validity is independent of chain integrity. |
signatures_verified | Count of offline-verified signatures: N Merkle-root signatures + 1 manifest signature (N+1). 0 for v1.0 packs (no embedded public key) — the tool emits a signatures not verifiable offline (v1.0 pack) note instead of failing. |
Trust model
The v1.1 pack carries its own SPKI public key. The signature proves the pack was not altered after signing by the holder of that key — it is trust-on-first-use. To pin trust to a specific signer out-of-band, an auditor can cross-check signature.signing_cmk_arn against AWS KMS GetPublicKey and confirm the embedded SPKI key matches. Malformed pack JSON returns isError: true (error: "invalid_pack_json") — the tool never throws past its boundary.
This tool imports only @hookwarden/canonical-json (the shared RFC 8785 byte-equality anchor) + node:crypto. It does not load the engine or rule pack and does not run the drift gate.
Integrity claims
- Zero outbound TCP — the published tarball passes
pnpm inspect-bundle, a structural CI gate that rejects any source-file import of HTTP clients (node:http,axios,undici, etc.) and any direct dependency outside the documented allowlist. - Drift detection — every
scan_handlercall cross-checks the runtime-resolved engine + rule-pack hashes against the build-timedist/build-manifest.json. Mismatch returnsisError: truewith structuredengine_drift/rules_driftpayload. No env-var opt-out. See drift-detection — including the v0.8.0 engine-content-hash limitation. - No telemetry — the tarball declares no analytics or tracking SDKs (
@datadog/*,@segment/*,@sentry/*, etc.) independenciesor transitively in source-file imports. - Open source — Apache-2.0 at
Hookwarden/hookwarden/packages/mcp/. Engine + rule pack land in the same monorepo and ship under the same license.
Links
- Documentation — getting started, tool surface reference, drift detection
- npm package
- GitHub repository
- Apache-2.0 License
Copyright 2026 Hookwarden contributors. Apache License 2.0.
