About Hookwarden — Webhook Integrity
Hookwarden exposes deterministic webhook-signature checks and evidence-pack integrity verification to coding agents. Its checks run locally without network access and return structured findings.
Key features
- 01Scan webhook handlers for signature checks
- 02Report verified, unverified, or review verdicts
- 03Verify evidence-pack hash chains offline
- 04Check embedded Merkle-root signatures offline
- 05Support 27 webhook providers
Use cases
- 01Audit a handler for missing signature verification
- 02Check whether an evidence pack was altered
- 03Review webhook code across supported providers
MCP support
The MCP server provides `scan_handler` for inspecting webhook code and `verify_audit_chain` for offline evidence-pack verification. It runs locally with no network egress or authentication.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
2Saved tool list · updated 2w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| code | string | No | |
| files | object | No | |
| language | enum | No | js | ts | python | php | go |
| provider | string | No |
Details
From the project README
The first MCP server doing webhook signature verification.
Local. Deterministic. Zero network. JS / TS / Python / PHP / Go — 27 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 23 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.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.