Run a FlareDeck agent on your own machine: it picks up cards assigned to it, works each one in a checkout of its own, and reports back on the card.
Flagged as AI-agent capability risk
Allowed by default with warning: agent-facing configuration or capability changes need review before use.
AI security review
scanned · by lpm-firewall-ai
The package enables manager-issued shell execution on the local machine by default, outside the agent's documented card workflow and shell restrictions. The local opt-out and FlareDeck authorization mitigate the risk, but the undisclosed remote capability warrants a warning.
Static reason
One or more suspicious static signals were detected.; previous stored version diff introduced dangerous source
Trigger
The running agent receives a manager command event from FlareDeck.
Impact
A workspace manager can execute commands as the local agent user. This is a powerful, insufficiently disclosed capability that warrants a warning.
Mechanism
It claims the command with the agent's authorization, runs it through bash in the configured working directory, and posts the result to FlareDeck. A local remote-off setting prevents execution.
Evidence
dist/serve.jsdist/remote-exec.jsREADME.mdConfigured working directoryFlareDeck agent config directory/remote-off
Network endpoints1
Configured FlareDeck MCP origin (mcpUrl)
Decision evidence
public snapshot
Evidence for warning
When the agent receives a manager command event, it calls the remote command handler.
The handler runs the claimed command through bash in the agent's working directory and sends its output back to FlareDeck.
Remote commands are enabled unless a local remote-off file exists; users can disable them with a CLI command.
The README describes the agent's work and enforced shell restrictions but does not mention manager-issued remote shell commands.
Evidence against
Commands are claimed from FlareDeck using the agent's own authorization, and the handler can be disabled locally.
The package is designed to run a FlareDeck agent on the user's machine and report work to its workspace.
Behavioral surface
SourceChild process · Crypto · Environment variables · Eval · Filesystem · Network · Shell · Web socket
Supply chainHigh entropy strings · URL strings
ManifestNo license
Scanned 132 files · 1.62 MB of source
14 external domains referenced
127.0.0.1
api.github.com
app.flaredeck.app
flaredeck-agent-861259873190.us-central1.run.app
generativelanguage.googleapis.com
github.com
jam.dev
mcp.cnjure.design
mcp.flaredeck.app
mcp.jam.dev
mcp.sceneflo.ai
openrouter.ai
pypi.org
registry.npmjs.org
Findings
1 critical · 6 high · 5 medium · 8 low
Where they are, most severe file first. Code opens a file's flagged lines, with their evidence, below.
382 { type: 'text', text: `This is the screen at ${viewport.width}×${viewport.height}. Judge only what is plain to see; never give a size in pixels.${takeStuckNote()}` },
// npx puts npm[redacted] in the environment, and the environment beats the config FILE
32// we just wrote — which is why `npm config set prefix` appeared to work and the install
⋯
58functionshellProfile() {
59const home = homedir();
60const shell = process.env['SHELL'] ?? '';
61const candidates = shell.includes('zsh')
62 ? ['.zshrc', '.zprofile', '.profile']
63 : ['.bashrc', '.bash_profile', '.profile'];
'never'
35});
⋯
44}
45/** `https://github.com/Owner/Repo(.git)` → `owner__repo`: the clone's folder name. */