More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    mcp-bastion: Self-Healing, Secure MCP Server Proxy

    1

    Keeps your MCP servers connected and blocks tool poisoning and rug-pull attacks, with no changes to your servers.

    Free

    1 useSecurity scanned
    mcp-bastion: Self-Healing, Secure MCP Server Proxy

    mcp-bastion: Self-Healing, Secure MCP Server Proxy

    Example session with this skill installed

    My filesystem MCP server keeps dropping in the middle of long sessions, and then the agent says "No such tool available". Put mcp-bastion in front of it so it reconnects on its own.

    • Read your context and instructions
    • Compiled the mcp-bastion

    I'll put mcp-bastion between your client and the filesystem server. Your server runs unchanged, and you can undo this by restoring your old config.

    1. Move the server into bastion.config.json

    I copied your existing filesystem entry as it was:

    {
      "servers": {
        "filesystem": {
          "command": "npx",
          "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects"]
        }
      },
      "reconnect": { "auto": true },
      "healthCheck": { "enabled": true }
    }
    

    2. Point your client at Bastion

    Your .mcp.json now has one entry. I used an absolute path, because clients often start servers from a different folder:

    {
      "mcpServers": {
        "bastion": {
          "command": "npx",
          "args": ["-y", "mcp-bastion", "--config", "/Users/you/bastion.config.json"]
        }
      }
    }
    

    3. Check it's working

    After the restart I called bastion__status:

    {
      "servers": [
        {
          "name": "filesystem",
          "state": "connected",
          "connected": true,
          "transport": "stdio",
          "authenticated": null,
          "tools": 14,
          "lastError": null
        }
      ]
    }
    
    WhatResult
    Filesystem serverConnected, 14 tools
    Tool namesNow prefixed with the server name, e.g. filesystem__read_text_file
    Bastion control toolsbastion__status, bastion__reconnect, bastion__security, bastion__compliance
    SecurityOn, balanced profile (blocks rug pulls and injection, warns on heuristics)
    Test readfilesystem__read_text_file returned the file through Bastion

    What changes for you: if the server drops again, Bastion reconnects it automatically. If a tool call fails mid-session, I'll check bastion__status, call bastion__reconnect once, and retry. If it still fails, I'll show you the server's last error instead of guessing.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    The problem

    When an MCP server crashes or disconnects mid-session, your agent just sees "No such tool available". It can't tell a dropped server from a tool that never existed, and it can't reconnect. You have to restart things by hand.

    Third-party MCP servers bring a second risk. A server can look harmless when you install it, then change a tool's definition later, or hide instructions in a tool description that steer your agent.

    What this skill does

    This skill teaches your agent to set up and use mcp-bastion, an open-source proxy that sits between your AI client and your MCP servers.

    • Self-healing connections. Bastion health-checks every server and reconnects automatically with backoff. Your agent gets bastion__status and bastion__reconnect so it can find and fix a dropped server itself.
    • Runtime security, on by default. Bastion pins every tool definition and blocks a tool that changes after it was pinned (a "rug pull"). It also flags poisoned tool descriptions, injection in tool arguments, tool shadowing across servers and leaked secrets in results.
    • Audit trail (opt-in). Bastion records a structured event for every tool call, mapped to the NIST AI RMF and the OWASP LLM Top 10, and sends it to a file, a webhook or OpenTelemetry.
    • No changes to your servers. You move your existing server entries into one config file. To undo, restore your old client config.

    The skill also tells the agent how to behave: reconnect once and report if it still fails, never work around a security block, and leave clearing a blocked tool to you.

    Works with

    Claude Code, Cursor, Claude Desktop, Cline, Windsurf and any MCP client. The skill includes the config file location for each one.

    Requirements

    • Node.js 20 or newer
    • At least one MCP server you already use

    Free and open source

    mcp-bastion is free under the Apache-2.0 licence. Its security coverage is measured on the open mcp-defense-bench benchmark.

    Created and maintained by Gowthaman Arumugam.

    Requires Node.js 20+. Agent cannot bypass rug pull blocks. Absolute paths required for config files. Best-effort redaction for sensitive arguments.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    Node.js 20+
    Bastion MCP
    MCP server
    Node.js

    Permissions

    Terminal / Shell
    Network Access
    Read Files
    Write Files

    File Scopes

    **/bastion.config.json
    .mcp.json
    .cursor/mcp.json
    ~/.cursor/mcp.json
    **/claude_desktop_config.json
    **/cline_mcp_settings.json
    ~/.codeium/windsurf/mcp_config.json

    Terminal: runs npx to start mcp-bastion, or `claude mcp add` in Claude Code. Read and write files: reads your existing MCP client config, creates bastion.config.json, and updates the client config to point at Bastion (it backs up the original first and shows you the changes). Network: npx downloads the mcp-bastion package from npm; Bastion only contacts the MCP servers you configure. The skill copies existing server entries as they are and never prints secret values.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    It sits between your AI client and your MCP servers. It health-checks each server and reconnects crashed local servers automatically with backoff. It also gives your agent bastion__status and bastion__reconnect, so it can check a dropped server and recover it instead of failing with “No such tool available”.

    Any MCP client, including Claude Code, Cursor, Claude Desktop, Cline, Windsurf and VS Code Copilot. You need Node.js 20 or newer to run mcp-bastion.

    The agent copies your existing server entries, unchanged, into a bastion.config.json file. It then replaces them in your client config with one bastion entry that runs npx -y mcp-bastion --config <path>. After you restart the client, it checks that every server shows as connected with bastion__status. Reconnection is automatic. To undo, restore your original client config.

    There are three profiles. observe only warns and never blocks, which is useful for a trial week. balanced is the default: it blocks high-confidence attacks such as rug pulls, argument and command injection, and cross-server exfiltration, and warns on heuristic findings such as poisoned descriptions. strict blocks on every finding, including the heuristic ones.

    This skill is free. You get a SKILL.md that teaches your agent when to use mcp-bastion, how to set it up for your client, and how to respond when a server drops or a call is blocked. mcp-bastion itself is free and open source (Apache-2.0). npx downloads it from npm the first time it runs.

    No, by design. Clearing a rug-pull block is a security decision, so the agent has no tool for it. Otherwise a manipulated agent could approve the very tool it was blocked from. You review the changed tool definition yourself. If it’s legitimate, restart Bastion and it re-pins the current definitions.

    New in LLM Ops