All articles
Open Source

Helixar MCP 0.8.0 Adds Pre-Install MCP Security Checks

The open-source update reviews tool metadata across MCP servers, checks OAuth discovery, and verifies signed Human Delegation Provenance records.

Three tool inputs meet at a central review point

An MCP client can read a tool’s name, description, and input schema before making a single call. Those fields help an agent decide which tool to use. That makes the manifest part of the security review when a team connects a Model Context Protocol (MCP) server. [2]

Helixar Labs has published Helixar MCP 0.8.0 on GitHub. The release gives developers and security teams three ways to review an MCP integration. They can inspect the tool metadata available to an agent, check OAuth discovery on an HTTPS endpoint, and verify a signed record of delegation against a trusted issuer key. [9]

What does the MCP manifest scanner inspect?

The scanner takes a manifest as JSON or a URL. Quick mode runs eight Sentinel rules. Deep mode runs 32 Sentinel rules and adds composition checks. Six of the Sentinel rules look for tool poisoning indicators, including invisible Unicode, instructions inside descriptions or schemas, and annotations that appear to conflict with the operations a tool describes. Findings include remediation and references to the relevant categories in OWASP and MITRE ATLAS. [1] [7] [8]

An annotation such as readOnlyHint is a declaration by the server. A reviewer can compare it with the documented operation and the permissions the team is about to grant. A finding identifies a specific inconsistency to investigate. The decision still depends on what the tool does in its deployed configuration. [2]

A scan can cover two to ten servers, including up to three fetched manifests. Consider an agent connected to one server for customer records, another for reading public pages, and a third for sending notifications. Each server could pass a review of its own metadata. The combined scan draws attention to all three capabilities appearing in the same agent session. It also flags tool-name collisions and descriptions that direct the agent to use a tool from another scanned server. [1] [3]

  • Customer records
    Private data
  • Public pages
    Untrusted content
  • Notifications
    Outbound channel
One agent session
Three servers can give one agent access to private data, untrusted content, and an outbound channel. The composition check raises the combination for review.

A useful review starts with the servers that a particular agent will actually load. That scope matters because the same manifest can produce a different security decision in two environments. A team that keeps outbound tools out of a data-review session may accept a combination that would deserve more scrutiny in a general-purpose assistant.

These findings come from the names, descriptions, and schemas a server declares. Reviewers can use them to decide which permissions and network destinations to inspect in the actual deployment.

What does the OAuth configuration check cover?

The auth probe reports eight checks covering the 401 challenge, protected-resource and authorization-server metadata, PKCE S256, resource indicators, dynamic client registration, CORS, and HTTPS. It uses GET and OPTIONS within a ten-second overall deadline. The CORS result describes what the preflight advertises. The endpoint’s handling of Origin on a real MCP request still needs to be confirmed separately. [1] [4]

OAuth discovery also crosses a boundary of its own. The check starts from the target origin and can follow the first advertised authorization server to its discovery metadata. It evaluates the advertised configuration without sending credentials or invoking the authorization or token endpoint. That gives reviewers a way to catch a missing discovery document before they test a full OAuth flow. [4]

What does HDP verification establish?

Human Delegation Provenance (HDP) records the chain linking a human principal to delegated agent actions. Given an HDP token and the issuer’s Ed25519 public key, the verifier checks signature and chain integrity offline and reports the recording period. An optional Mermaid diagram shows the hops in order. Each caller is responsible for choosing a trusted issuer key and applying its own access policy. [1] [5]

Integrity and timing have separate meanings in the report. A signature can verify even when a hop was recorded after the token’s declared recording period. Keeping those results separate helps an investigator interpret a retained record and decide whether the timing needs a closer look. [5]

The same verifier can run locally or on Cloudflare Workers. HDP SDK 0.2.3 compiles its token schema validator during the build, which lets the Worker load it with runtime code generation disabled. [6]

Use Helixar MCP 0.8.0

MCP clients can read the rule catalogue at helixar://rules and inspect individual rules for evidence and remediation guidance. The pre-install review prompt picks a quick scan when no key is supplied. On an HTTPS target, it also asks the client to check authentication configuration. That gives reviewers a repeatable starting point across different servers. [1]

Connect an MCP client to https://mcp.helixar.ai/mcp over Streamable HTTP to use the four hosted tools. A local stdio installation adds ReleaseGuard as a fifth tool. For the local setup, check out the Apache 2.0 source release tagged v0.8.0 and follow the stdio instructions in the repository. The merged pull request records the test results. [9] [1]

Get Helixar MCP 0.8.0 on GitHub