Saltar al contenido
MCP ThesaurusMCP Thesaurus

Dredd MCP

ComunidadIncomplete39/100Reclamar

streamable-httpupdated 4mo ago

Dredd MCP is a pre-invocation security check for the Model Context Protocol ecosystem. Before your agent calls a tool on any other MCP server, Dredd renders a verdict: BLOCK, ADVISORY, or ALLOW. Every verdict is HMAC-signed and cites the IOC or behavioral signal that drove the decision.

CódigoWeb

¿Qué puedes hacer con Dredd MCP?

Dredd MCP — Pre-Flight Security for the MCP Ecosystem

"Jeevesus saves. Dredd judges."

Dredd MCP is a pre-invocation security check for the Model Context Protocol ecosystem. Before your agent calls a tool on any other MCP server, Dredd renders a verdict: BLOCK, ADVISORY, or ALLOW. Every verdict is HMAC-signed and cites the IOC or behavioral signal that drove the decision.

The MCP ecosystem has had no defender. Three PyPI ML packages were compromised in eight days during late April 2026. Twenty-plus MCP-named GitHub repositories were caught serving SmartLoader malware in the wild. The official MCP Registry was clean of those when we measured — but the typosquat surface is wide open.

Dredd is the layer that catches the next compromise before the malicious tool gets called.

What's New — Dependency-Graph Checking (Shai-Hulud class)

Dredd no longer stops at the named server. It walks the transitive dependency graph.

The Shai-Hulud worm taught the ecosystem the hard lesson: the malicious code is rarely in the package you installed — it's three levels down, in a transitive dependency that got its publish token stolen. A check that only vets the server you named is blind to exactly the attack that's been hitting npm and PyPI.

So every preflight now resolves the target server's full npm/pypi dependency graph and joins every transitive package against our continuously updated IOC corpus — including the OSV malicious-package feeds for npm and PyPI. If a known-malicious package is buried anywhere in the tree, Dredd blocks the call before the tool runs.

The verdict is signed (HMAC-SHA256) and now carries a dep_graph field telling you whether transitive supply-chain risk was actually evaluated for that target:

{
  "verdict": "BLOCK",
  "severity": "critical",
  "dep_graph": { "evaluated": true, "packages_checked": 247, "malicious_transitive": 1 },
  "signature": "sha256=..."
}

If the server doesn't expose a resolvable manifest, dep_graph.evaluated is false and Dredd drops to the advisory tier — it tells you it couldn't see the tree rather than pretending it's clean.

What Dredd Checks

Every preflight call evaluates these signals:

  1. Compromised dependency — including transitive (Shai-Hulud class). The target server's package manifest is parsed, its full npm/pypi dependency graph is resolved, and every package — direct and transitive — is joined against our continuously updated IOC corpus (Socket, Aikido, GitGuardian, ReversingLabs, Phylum, StepSecurity, Wiz, plus OSV malicious-package feeds for npm and PyPI). If the server, or anything it pulls in transitively, pins lightning==2.6.2 or any other known-compromised version, the call is blocked. The dep_graph field on the verdict reports whether the transitive tree was evaluated.
  2. Tool surface drift. The list of tools the server exposes today versus the snapshot the user originally approved. New tools that appeared since the last review trigger an advisory. Mid-session rugpull is the threat model.
  3. Remote URL drift. The server's runtime endpoint compared against the URL it published in the registry. A server quietly calling out to a different host than the one you signed up for is a hijack signature.
  4. Permission escalation. A server requesting write or exec permissions it did not have last week.

The verdict comes back signed in under 200 ms (Cloudflare-edge cached, 5-minute TTL). The hook fails open by default — if our endpoint is ever down, Dredd does not brick your tooling.

Install — Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "dredd": {
      "url": "https://analytics.dugganusa.com/api/v1/dredd/mcp"
    }
  }
}

Restart Claude Desktop. You'll see Dredd available with one tool: check_mcp_server.

Install — Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "dredd": {
      "url": "https://analytics.dugganusa.com/api/v1/dredd/mcp"
    }
  }
}

Test from terminal

curl -X POST https://analytics.dugganusa.com/api/v1/dredd/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

You should see one tool: check_mcp_server.

The Tool

check_mcp_server

Argument Type Required Description
server string yes MCP server name (e.g. io.github.foo/bar) or substring
version string no Optional semver of the server
tool string no Optional name of the specific tool being invoked

Returns a JSON verdict:

{
  "success": true,
  "server": "io.github.foo/bar",
  "version": "1.2.3",
  "tool": "list_files",
  "verdict": "ALLOW",
  "severity": "clean",
  "findings_count": 0,
  "findings": [],
  "dep_graph": { "evaluated": true, "packages_checked": 247, "malicious_transitive": 0 },
  "checked_at": "2026-05-04T20:00:00Z",
  "ttl_seconds": 300,
  "signature": "sha256=..."
}

Verdict values:

Verdict Severity tier What to do
BLOCK critical or high Refuse the invocation. Do not call the target tool.
ADVISORY medium or advisory Surface to the user; let them decide.
ALLOW clean Proceed.

The signature field is an HMAC-SHA256 of the canonical verdict body using a server-side secret. Hooks should verify this to defeat MITM-forged "all clear" responses (verification key distributed out of band on request).

The Public Watchtower

Real-time aggregate dashboard of every active finding across the registry — free, no auth, no email gate:

https://analytics.dugganusa.com/api/v1/dredd/watchtower.json

Returns counts by severity, recent findings, current verdict (CLEAN / WATCH / HIGH / CRITICAL).

Updated continuously as the daily fetcher + correlator pipeline runs against the registered MCP corpus.

How Dredd Works

       ┌─────────────────────────────────────────────────┐
       │  Your Agent (Claude Desktop / Cursor / custom)   │
       │     calls check_mcp_server(server, version, tool)│
       └─────────────────┬───────────────────────────────┘
                         │ JSON-RPC over HTTPS
                         ▼
        ┌────────────────────────────────────────────────┐
        │  Dredd MCP — analytics.dugganusa.com/api/v1/dredd/mcp │
        │  - look up findings for (server, version, tool)│
        │  - aggregate severity, render verdict          │
        │  - HMAC-sign canonical verdict                 │
        └────────────────┬───────────────────────────────┘
                         │
                         ▼
        ┌────────────────────────────────────────────────┐
        │  mcp_findings index — populated by             │
        │  daily fetcher + correlator joining the        │
        │  resolved transitive dep graph × IOC corpus    │
        │  (Socket, Aikido, GitGuardian, ReversingLabs,  │
        │   + OSV malicious-package feeds: npm & PyPI)    │
        └────────────────────────────────────────────────┘

The correlation cadence today is 12 hours (08:30 UTC and 20:30 UTC). When a real compromise lands in the registered-MCP corpus, cadence tightens.

Trust Posture

  • HMAC-signed responses. Hook implementations should verify the signature field on every verdict.
  • Fail-open by default. If our endpoint is down, Dredd does not brick your tooling — it returns "advisory: backend unavailable" and lets the user decide. Document override (DREDD_BYPASS=<reason>) for critical workflows.
  • Read-only. Dredd never modifies your environment. Verdict only.
  • No tool argument leakage. Hooks should send (server, version, tool) only — never the contents of tool arguments. Those stay on your machine.
  • Validated corpus, not just a big one. The IOC corpus behind every verdict is independently checkable on four live, no-auth endpoints: novelty (/api/v1/feed-uniqueness, ~75%+ of our IOCs aren't in ThreatFox), timeliness (/api/v1/kev-lead, live kev-lead ledger vs CISA KEV), accuracy (/api/v1/spamhaus-validation), and liveness (/api/v1/feed-efficacy, opt-in consumer reports of when our indicators actually fire on real traffic — proof the feed is operationally live, not just large).
  • 95% epistemic ceiling. We cap our claims at 95% per DugganUSA's epistemic humility rule. Coverage gap: about 60-70% of MCP servers in the registry today don't expose a public source repository, which means Dredd cannot inspect their dependency tree. The advisory tier exists for those.

The Family

Dredd is the 13th member of the DugganUSA defender family — and the first MCP-native member:

Companion MCP server: Jeevesus — natural-language threat intelligence search across 38M documents. Jeevesus saves. Dredd judges.

License

MIT — see LICENSE.

Support

Built in Minneapolis. Defender-grade. Read-only. Receipts do the work.

DugganUSA Defender Family

Same threat corpus, surfaced wherever you live. Open source, MIT licensed, receipts on every repo.

Plugin Surface
dugganusa-scanner-core Core IOC scanning engine
dugganusa-vscode VS Code extension
dugganusa-splunk Splunk Technology Add-on
dugganusa-slack Slack bot
dugganusa-raycast Raycast extension
dugganusa-sentinel Microsoft Sentinel TAXII connector
dugganusa-obsidian Obsidian plugin
dugganusa-nvim Neovim plugin
dugganusa-elastic Elastic / OpenSearch integration
dugganusa-edge-shield Cloudflare Worker
dugganusa-cli CLI scanner
dugganusa-chrome Chrome extension
dugganusa-action GitHub Action
dredd-mcp (this repo) Pre-flight MCP security (this repo)

Backed by the live DugganUSA threat intel platform: analytics.dugganusa.com.

Jeevesus saves. Dredd judges.