agenthub-multiagent-mcp@1.72.0
MCP server for AgentHub multi-agent communication
AI Security Review
scanned 6d ago · by lpm-firewall-aiThe MCP server automatically writes global Claude Code skills and hooks, including a session-start prompt that directs the agent to process inbox messages before the user’s request. This is a meaningful agent-control capability exposed by default, though its stated purpose and side-effect limits do not establish malicious intent.
Static reason
No blocking static signals were detected.
Trigger
Starting the package’s MCP server; the hooks are enabled by default and can be disabled with their respective environment settings.
Impact
This gives the package influence over later Claude Code sessions and can cause the agent to read and answer team messages before the user asks it to do so.
Mechanism
The server writes bundled skills and hook entries into ~/.claude, including a SessionStart hook that injects instructions to check and respond to AgentHub messages.
AI rationale
The MCP server automatically writes global Claude Code skills and hooks, including a session-start prompt that directs the agent to process inbox messages before the user’s request. This is a meaningful agent-control capability exposed by default, though its stated purpose and side-effect limits do not establish malicious intent.
Evidence
dist/index.jsdist/inboxHookInstall.jsdist/inboxHookCli.jsdist/activityHookInstall.jsdist/skillInstaller.js~/.claude/settings.json~/.claude/commands/~/.claude/skills/<project>/.agenthub/activity
Network endpoints1
agenthub.contetial.com
Decision evidence
public snapshotAI called this Suspicious at 88.0% confidence as Unknown with medium false-positive risk.
Evidence for warning
- The package’s main entrypoint starts the MCP server and automatically installs bundled skills into the user’s ~/.claude directory unless opted out.
- On MCP startup, the package installs an inbox hook into ~/.claude/settings.json by default; AGENTHUB_AUTO_INBOX=false, 0, or off disables it.
- The hook prompts Claude to check the AgentHub inbox and handle messages before the user’s first request, including replying through AgentHub MCP tools.
- The package also installs a PostToolUse activity hook into ~/.claude/settings.json by default unless AGENTHUB_AUTO_ACTIVITY=false.
- The worker sends its API key to the configured AgentHub WebSocket server, whose default host is agenthub.contetial.com.
Evidence against
- The inbox hook’s prompt says messages are input rather than instructions and tells the agent to stop before side effects such as deploying, publishing, deleting, or spending.
- The worker’s credential is sent to the configured AgentHub service as its authentication key.
- The shell startup hook edits shell configuration only when the user explicitly runs agenthub-setup install-shell.
- The manifest declares no lifecycle scripts.
Source & flagged code
6 flaggeddist/commands/setup-shell.jsView file
25Trigger-reachable persistence chain: manifest.bin -> dist/setup.js -> dist/commands/setup-shell.js
L25: export function detectShell() {
L26: const shellPath = process.env.SHELL;
L27: if (!shellPath)
...
L35: export function rcFilePath(shell) {
L36: const home = os.homedir();
L37: return path.join(home, shell === "bash" ? ".bashrc" : ".zshrc");
L38: }
High
Trigger Reachable Persistence
A manifest entrypoint or package-local install chain reaches persistence behavior.
dist/commands/setup-shell.jsView on unpkg · L2525export function detectShell() {
L26: const shellPath = process.env.SHELL;
L27: if (!shellPath)
...
L35: export function rcFilePath(shell) {
L36: const home = os.homedir();
L37: return path.join(home, shell === "bash" ? ".bashrc" : ".zshrc");
L38: }
Medium
Install Persistence
Source writes installer persistence such as shell profile or service configuration.
dist/commands/setup-shell.jsView on unpkg · L25dist/worker.jsView file
Findings
3 High6 Medium4 Low
HighTrigger Reachable External Ai Agent Control Surface Mutationdist/worker.js
HighTrigger Reachable Persistencedist/commands/setup-shell.js
HighSemantic Analysis Limiteddist/brain/memvidBrain.js#virtual:normalized:round1
MediumNetwork
MediumEnvironment Vars
MediumInstall Persistencedist/commands/setup-shell.js
MediumShips Native Binarynative/agenthub-memvid/agenthub-memvid.win32-x64-msvc.node
MediumShips Build Helperdist/run-claude.sh
MediumStructural Risk Force Deep Review
LowScripts Present
LowFilesystem
LowHigh Entropy Strings
LowUrl Strings
Affected versions and remediation
This report applies to agenthub-multiagent-mcp@1.72.0.
See all versions of agenthub-multiagent-mcp for other recorded verdicts.
- Review the evidence and your use of agenthub-multiagent-mcp@1.72.0 before allowing it. Restrict the permissions described in this report.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.