mecagents
Read-only agents with code-checked findings
Four example agents built on mecbots: a local model with read-only MCP tools, and code that verifies every finding.
About
mecagents is an add-on to mecbots with four example agents. An agent is the riskier tier above a bot: a local model (qwen3-coder-next, chosen from the mechub bake-off) runs once via Hermes with fwskillsshare skills and only read-only MCP tools. Tokens are read-only and scoped at the MCP server.
Code, not the model, decides. Code fetches the data independently, verifies each finding and grades the run from parsed structured findings; run records are scrubbed when saved. Measured reliability of the local model is variable: it sometimes tries tools it is not offered (refused and flagged), and verdicts can differ between runs, which is why code verifies every finding.
Discord is not sovereign: it is a cloud service, used only as the exit point for scrubbed reports. A self-hosted push option (ntfy) is planned.
Features
- pcicomplyagent: PCI DSS firewall check, also triggered by a device commit through a Junos event script
- bpcheckagent: best-practice audit
- driftagent: config drift against a golden config
- socagent: investigates items socbot escalates
- Read-only, server-scoped tokens; the model is offered no other tools
- Alpha: results vary run to run
Quick start
Clone mecbots, then overlay mecagents, then follow docs/SETUP.md for the read-only token and Hermes profile
git clone https://github.com/mechubsec/mecbots ~/mecbots
git clone https://github.com/mechubsec/mecagents /tmp/mecagents
cp -r /tmp/mecagents/bots /tmp/mecagents/lib /tmp/mecagents/tests /tmp/mecagents/systemd ~/mecbots/