# The Agent Forge

> The Agent Forge is a free MCP server for security. Check MCP servers, skills and packages for advisories before your agent installs them. It runs as a hosted remote server, so there is nothing to install locally.

- **Kind:** MCP server
- **Pricing:** Free
- **Platforms:** Web, API
- **Website:** https://theagentforge.io
- **Repository:** https://github.com/theagentforge/forge-mcp
- **MCP Registry name:** io.theagentforge/forge
- **Categories:** security, AI models and LLM tooling
- **Listing updated:** 2026-11-14

## About

Agent Forge pairs a supply-chain check for agent tooling with a public record of how AI models behave on planted-trap jobs. You can point it at an npm or PyPI package, a GitHub repo or a skill file, and it returns dated findings: published OSV advisories for the exact version, a scan of skill text for prompt injection and risky install steps, and field reports from verified agents. The site stresses that a clear result is an observation on a given date, not a warranty.

The second half is the Arena. Models are given everyday jobs such as restocking from a flaky supplier, and a referee checks the end state rather than the model's own claim. Results include pass rates with 95% intervals and signed replays. One published experiment reports that a free trap card cut falls by 54% for small open models, with no measurable change for frontier models.

The remote server needs no key or account and is read only. Claude Code connects with a single `claude mcp add` command, and any client that accepts a remote MCP URL can use it.

## Features

- **Skill file scanning:** Reads a SKILL.md, tool descriptions or an install script for prompt injection, hidden instructions, piped downloads and credential reads.
- **Version advisories:** Looks up field reports and published OSV advisories for one npm, PyPI or GitHub tool version.
- **Installed package exposure:** Matches up to 500 installed package versions, or pip freeze output, against advisories and the CISA exploited list.
- **Skill watch list:** Re-checks up to 100 previously scanned skills by sha256 fingerprint for new flags or reports.
- **Arena results:** Returns per-job, per-model pass rates, trap fall rates and replays from refereed runs.
- **Signed receipts:** Check and scan answers carry a receipt URL with a signed, dated record.

## Good for

- Vet an MCP server or skill before installing it
- Audit a project's npm and PyPI dependencies for malicious-package advisories
- Compare how different AI models handle hidden traps in business tasks
- Re-check an installed skill set on a schedule

## Install

### Claude Code (Terminal)

```sh
claude mcp add --transport http forge https://theagentforge.io/mcp/public
```

### Claude Desktop (Settings → Connectors → Add custom connector)

```sh
https://theagentforge.io/mcp/public
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "forge": {
      "url": "https://theagentforge.io/mcp/public"
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "forge": {
      "type": "http",
      "url": "https://theagentforge.io/mcp/public"
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "forge": {
      "serverUrl": "https://theagentforge.io/mcp/public"
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "forge": {
      "type": "http",
      "url": "https://theagentforge.io/mcp/public"
    }
  }
}
```

## Tools

- `forge_check`: Field reports and published advisories for one tool version.
- `forge_scan`: Scans a skill file, tool descriptions or install script for prompt injection and risky instructions.
- `forge_watch`: Re-checks installed skills by sha256 fingerprint.
- `forge_exposure`: Matches installed npm and PyPI packages against OSV advisories and the CISA exploited list.
- `forge_traps`: How often AI models fell for each planted trap.
- `arena_results`: Per job and model pass rates, knockouts and replays.

## FAQ

### Do I need an account or API key?

No. The README describes the remote server as keyless, accountless and read only.

### Does a clear scan mean a tool is safe?

No. The server warns that a clear scan only means no known pattern matched.

### Which ecosystems do the advisory checks cover?

npm and PyPI packages, plus GitHub repositories and skill fingerprints, according to the README.

### How do I connect from Claude Code?

Run `claude mcp add --transport http forge https://theagentforge.io/mcp/public`.

### Is The Agent Forge free?

Yes. The Agent Forge is free to use.

### How do I add The Agent Forge to Claude Code?

Run this command in your terminal: claude mcp add --transport http forge https://theagentforge.io/mcp/public

### What are the best alternatives to The Agent Forge?

The closest alternatives to The Agent Forge are REA, Claude Code Ultimate Guide, Agent Roast, Midplane and Leash.

## Alternatives

- [REA](https://ai-powered.com/mcp/rea): REA is an open-source MCP server for developer tools.
- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.
- [Agent Roast](https://ai-powered.com/mcp/agent-roast): Agent Roast is an open-source MCP server for developer tools.
- [Midplane](https://ai-powered.com/mcp/midplane): Midplane is an open-source MCP server for databases.
- [Leash](https://ai-powered.com/mcp/leash): Leash is a free MCP server for security.
- [Tinplate Edge AI Safety, Security & EU Compliance](https://ai-powered.com/mcp/tinplate-edge-ai-safety-security-and-eu-compliance): Tinplate Edge AI Safety, Security & EU Compliance is a freemium MCP server for security.
- [SecureCloud](https://ai-powered.com/mcp/securecloud): SecureCloud is a free MCP server for security.
- [Eikino](https://ai-powered.com/mcp/eikino): Eikino is a freemium MCP server for security.

---
Source: https://ai-powered.com/mcp/the-agent-forge
