Show HN: Conduct, open-source guardrails for LLM and MCP tool calls

Conduct is a governance/control plane for AI agents, letting one policy decide block / warn / audit / inject for every LLM call, shell tool, and MCP invocation before the action runs. The same policy applies to scheduled agents, developers running Cursor, and chat sessions. It ships with three surfaces: Guard (policy engine with signed config, hash-chained audit, fail-closed), Router (an LLM proxy that any Anthropic/OpenAI/Perplexity SDK can point at), and Lens (a chat surface that answers workspace questions, with every tool call also running through Guard).

The project contrasts itself with runtime firewalls like Straiker and Lakera: those report what an agent did after the fact, while Conduct decides what it can do before execution. Conduct claims coverage of LLM calls plus shell and MCP tools, fail-closed behavior, and an audit log with a SHA-256 hash chain. Audit integrity rests on signed workspace policy configs (tampered packs are rejected before enforcement) and a hash chain rooted at workspace genesis, so missing or altered entries break the chain.

Setup is claimed to take about 60 seconds: `pip install conduct-cli`, `conduct login`, `conduct sync`, after which every Claude Code, Cursor, Copilot, and Codex session on that machine is governed. Self-hosting is via `git clone https://github.com/sseshachala/conductai && docker compose up`, providing an API on localhost:8000 and a Canvas UI on localhost:3000. A free Discovery mode gives 14 days of read-only visibility into all AI actions, with the ability to promote a observed behavior into a blocking rule. The Router works by pointing any SDK to `https://api.conductai.ai/proxy/anthropic/v1/messages`, which runs every request through Guard before reaching the upstream provider.