# signet-eval

> signet-eval is an open-source MCP server for security. Deterministic policy rules that gate AI agent tool calls before they run. It is available on CLI, macOS, Linux and self-hosted.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** CLI, macOS, Linux, self-hosted
- **Repository:** https://github.com/jmcentire/signet-eval
- **MCP Registry name:** io.github.jmcentire/signet-eval
- **Categories:** [security](https://ai-powered.com/categories/security)
- **Listing updated:** 2026-11-18

## About

The README describes signet-eval as a local enforcement layer that sits between an AI agent and its tools. Each proposed tool call is checked against user-defined rules, and the result is allow, deny or ask. The authorization path uses regex, comparisons and vault lookups rather than an LLM, and advisory nudges are kept separate from those decisions.

It installs from crates.io or source and wires in through command hooks for Claude Code, Codex, Cursor, Antigravity and Muse Code, or runs as an MCP management server or an MCP proxy in front of upstream servers. Policies live in YAML, with locked self-protection rules that the agent cannot edit, an encrypted vault for credentials and spending limits, and optional masking of secrets read from shell env files.

The README frames it as a seatbelt rather than a sandbox: it runs with the same privileges as the agent, so the author recommends pairing it with OS-level isolation. It suits developers who run coding agents locally and want auditable, deterministic guardrails.

## Features

- **Deterministic rule evaluation:** Rules are matched in order with first-match-wins, using string, regex and numeric conditions instead of model judgment.
- **Host hook adapters:** Command-hook adapters cover Claude Code, Codex, Cursor, Antigravity and Muse Code.
- **Locked self-protection rules:** Built-in rules block the agent from editing policy files, the binary, or hook configuration, and cannot be removed through MCP.
- **Spending limits and vault:** An encrypted vault stores credentials and a spending ledger so purchase limits are checked across tool calls.
- **Secret masking:** Values from configured env files are replaced with key names in logs, reasons and displayed output.
- **MCP proxy and management:** A stdio proxy enforces policy in front of upstream MCP servers, and a management server edits rules conversationally.

## Good for

- Blocking destructive shell commands such as rm or mkfs during coding-agent sessions
- Requiring confirmation before force-pushes or other sensitive git operations
- Capping how much an agent can spend on purchases across a session
- Routing MCP tool calls through a policy checkpoint before they reach upstream servers
- Keeping API keys in local env files out of agent-visible output

## FAQ

### Does signet-eval use an LLM to make authorization decisions?

No. The README says the authorization path uses structured data only, such as regex, comparisons and vault queries, and advisory injection runs separately afterward.

### How is it installed?

The README lists cargo install signet-eval from crates.io or a source build from GitHub. It states there is no npm or PyPI package.

### Is it a complete sandbox for an adversarial agent?

No. The README says it is a seatbelt, not a cage, and that an agent with shell access at the same privilege level can evade string matching, so OS-level controls are recommended.

### Is signet-eval free?

Yes. signet-eval is open source, so you can use and self-host it for free.

### What are the best alternatives to signet-eval?

The closest alternatives to signet-eval are VITNA – Containment for rogue AI agents, agent-custody, Sanction, Two-Key and BarzelVault: AI Action Authorization Firewall.

## Where this comes from

- Description: written from the project's README and its MCP Registry entry.

## Alternatives

- [VITNA – Containment for rogue AI agents](https://ai-powered.com/mcp/vitna-containment-for-rogue-ai-agents): VITNA – Containment for rogue AI agents is an open-source MCP server for security.
- [agent-custody](https://ai-powered.com/mcp/agent-custody): agent-custody is a freemium MCP server for security.
- [Sanction](https://ai-powered.com/mcp/sanction): Sanction is a freemium MCP server for security.
- [Two-Key](https://ai-powered.com/mcp/two-key): Two-Key is an open-source MCP server for security.
- [BarzelVault: AI Action Authorization Firewall](https://ai-powered.com/mcp/barzelvault-ai-action-authorization-firewall): BarzelVault: AI Action Authorization Firewall is a paid MCP server for security.
- [mcpcut](https://ai-powered.com/mcp/mcpcut): mcpcut is an open-source MCP server for security.
- [sealkeep](https://ai-powered.com/mcp/sealkeep): sealkeep is an open-source MCP server for security.
- [skillsvault](https://ai-powered.com/mcp/skillsvault): skillsvault is a freemium MCP server for security.

---
Source: https://ai-powered.com/mcp/signet-eval
