@agentvault/mcp-server
The AI review rates the findings as likely false positive, but the risk score (100/100) still counts them.
Analysis record
- Analysed
- 3 weeks ago 27 October 2026
- Version
- v0.5.1
- Artifact
- SHA256 526de444…ebc0c8
- Detail
- Full list
Is @agentvault/mcp-server safe?
@agentvault/mcp-server connects an AI agent to AgentVault, a service for encrypted messaging, skill discovery, policy checks and audit logging. The package declares no permissions and no host permissions, and its outbound traffic goes to api.agentvault.chat, agentvault.chat and app.agentvault.chat, plus modelcontextprotocol.io, which is the MCP project's own documentation site.
The loudest findings are twelve hardcoded-token alerts, and every one is in a test file such as dist/tests/server.test.js or dist/tests/agent-teams.test.js. Those files hold throwaway strings used to stand up a test transport. They are not user credentials, and nothing reads them off a real machine. The genuine issue is packaging: the build ships its test suite inside dist/, so test literals end up in the published package.
The rest of the noise comes from scanner mechanics. Forty-three code-smell hits match ordinary JavaScript in a bundled file, and the endpoint list includes strings like result.trust and body.room that are property names in the code, not servers the package contacts.
A server that relays messages and writes audit logs does sit in the middle of an agent's tool calls, which is the one place to look hard. The scan found no hidden instructions in the tool descriptions and nothing that rewrites tool results before the agent sees them.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
YARA Rule Matches
8 rules| Severity | Rule | Hits | Files | Metadata |
|---|---|---|---|---|
| LOW | postinstall network communication | 13 | dist/server.jsdist/stdio.jsdist/server.d.ts +10 more | - |
| LOW | postinstall crypto operations | 10 | dist/index.d.tsdist/__tests__/agent-teams.test.jspackage.json +7 more | - |
| LOW | postinstall environment access | 1 | dist/stdio.d.ts | - |
| LOW | postinstall registry modification | 1 | server.json | - |
| LOW | postinstall obfuscation | 4 | dist/__tests__/phase2-edge-cases.test.jsdist/__tests__/server.test.jsdist/__tests__/agent-teams.test.js +1 more | - |
| LOW | postinstall system command | 3 | server.jsondist/__tests__/server.test.jsREADME.md | - |
| LOW | credential env files | 3 | dist/stdio.jsdist/__tests__/server-extended.test.jsdist/__tests__/server.test.js | - |
| LOW | postinstall file download | 6 | dist/__tests__/phase2-edge-cases.test.jsdist/__tests__/server-extended.test.jsdist/__tests__/server.test.js +3 more | - |
Finding Categories
YARA Rules Matched
8 rules(41 hits)MCP Server Analysis
MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.
AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
Tool poisoning is the first thing to check on any MCP server, and for @agentvault/mcp-server that category is empty. No hidden AI directives, no instruction tags in tool metadata, no invisible Unicode. The tools this package registers (encrypted messaging, skill discovery, policy checks, audit logging) are described for the human reading them, and nothing in the scanned data tells an agent to act silently.
The twelve critical findings are all MCP-TRANSPORT-HARDCODED-TOKEN, and all twelve sit in test files: dist/tests/server.test.js, dist/tests/server-extended.test.js, dist/tests/phase2-edge-cases.test.js and dist/tests/agent-teams.test.js. These are fixtures that stand up a stdio transport and hand it a literal string. The tokens belong to the publisher's test harness. What they expose is packaging: the build ships its test suite inside dist/, so test literals end up in the published tarball. That does not touch anything on the user's disk.
Credential scope is limited accordingly. No finding references .ssh, .aws/credentials, .kube/config or application_default_credentials.json. Nothing in the scanned data pairs a secret read with an outbound call to a domain outside the product.
The network picture holds together. The NET-FETCH hits in dist/server.js (lines 601 through 986) resolve against api.agentvault.chat, agentvault.chat and app.agentvault.chat, the service the package exists to front, plus modelcontextprotocol.io for the protocol spec. The endpoint list also carries entries like result.trust, profile.trust, body.room and agentvaultstandalonemcpserver.trust. Those have the shape of JavaScript property access inside bundled code, not registered hostnames, and they are the artifact the endpoint extractor produces on any minified file.
The strongest counterargument is architectural: a server that brokers encrypted messages and writes audit logs sits in the path of every tool call, which is exactly where result tampering would hide. That argument would matter if the evidence showed the software bending its own outputs. It doesn't. No interceptor or middleware file shows up among the findings, no tool result rewriting, and outbound traffic stays on AgentVault's own hosts. The 43 code-smell findings are low-severity pattern matches on ordinary bundled JavaScript, and the 18 IoC entries are dominated by the property-chain artifacts above.
One item deserves a check from the publisher. The test files shipped in dist/ contain hardcoded strings, and motiveflow should confirm those are fixtures and not live AgentVault credentials. If any are real, that is a leak of the vendor's own secrets and a reason to rotate them. It leaves the installed package harmless to the person running it.
Key Reasons
- No tool-poisoning findings at all: no hidden AI directives, instruction tags, or invisible Unicode in tool metadata
- All twelve critical MCP-TRANSPORT-HARDCODED-TOKEN alerts are in dist/tests/*.test.js fixtures, not production code
- Network calls in dist/server.js map to AgentVault's own hosts (api.agentvault.chat, agentvault.chat, app.agentvault.chat) and modelcontextprotocol.io
- No credential-access findings for .ssh, .aws/credentials, .kube/config, or application_default_credentials.json
- Remaining signal is 43 low-severity code-smell matches and IoC strings that are property access chains (result.trust, body.room)
False Positive Considerations
- Hardcoded-token rule matching stdio test fixtures shipped inside dist/tests
- IoC extractor reading JS property access chains like result.trust and body.room as hostnames
- Low-severity code-smell pattern matches on bundled JavaScript
- NET-FETCH hits on the vendor's own documented API domains
Reviewed 2026-10-30 · version 0.5.1; recommended action: monitor; model confidence 75%.
MCP version history
Risk trend by version
2 analyzed versions. Each point is the latest successful scan for that version; failed zero-score scans are hidden. Dates: First seen by Risky Plugins.
Pick any point on the chart, then open the Source tab to read that version's code.
Source Code Not Available
Source code is not available for this version of the extension.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace