mcpvessel
Tell Claude to add an MCP server. Claude cages it, runs it, watches everything it sends, and tells you if it's safe.
Potencialize seu agente
Instalação do publicador. Não é um endpoint do PlugMCP.
Host a copy with our keys and logs.
Ainda não há uma versão empacotada no registro. Siga o README do repositório.
URL remota ou um comando na sua máquina? Como as duas instalações diferem
Rodar no PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
AbrirRun mcpvessel init once, then tell Claude to add any MCP server, even one you would never trust. Claude installs it in a deny-default cage (secrets it cannot leak, no network it was not allowed), uses it on your behalf, surfaces everything it tries to send, and tells you plainly whether it is behaving or hiding something. You never vet a server or touch a cage. Nothing slips past you.
An MCP server runs as a subprocess with your full user permissions. The protocol does not sandbox it, so an installed server can read your SSH keys, cloud credentials, and .env files, run arbitrary commands on your machine, and send any of it anywhere.
This is not theoretical. CVE-2025-6514 (rated critical) is host remote code execution from connecting to an untrusted server, and audits keep finding thousands of vulnerable public servers. You cannot read the source of every server you want to try, and safe today does not mean safe after the next update.
mcpvessel makes that a thing you no longer have to weigh. Every server runs alone in its own container on an isolated network, its outbound traffic filtered by a gateway that opens only the hosts you allow, and your keys held outside the cage where it cannot reach them. So Claude can run a server you have never vetted, and the worst it can do is try. You run none of it: Claude installs, cages, and watches each server, and tells you the truth about what it does.
Install the binary, then run one command:
# macOS via the Homebrew cask (on Linux see Install).
brew install --cask okedeji/tap/mcpvessel
# One command sets up everything.
mcpvessel init
init is the whole setup. It provisions the runtime, installs the mcpvessel skill into Claude Code so Claude knows how to drive it, and installs mcpvessel's own docs as a caged server so Claude knows the tool inside out, all in one go. From here you do not run mcpvessel commands. Claude does.
Open a new Claude Code session and just ask:
you: add the GitHub MCP server so I can use it here.
Claude: Installed github-mcp-server into a cage and served it to this session.
Its network is deny-default, so I will hold anything it reaches that you
have not approved, and I am watching what it sends. Ask me to use it.
And when a server turns on you, as the notes server does above, Claude tells you before you think to ask.
You never touched a cage, an allow-list, or a config file. That is the point.
mcpvessel records every request each caged server makes, continuously, even the ones a server fires off while you are away from the keyboard. Claude reads that record and reports it: a host you did not expect, a secret leaving, a server reaching out during a call that had no business touching the network. When a granted secret shows up in an outbound request, mcpvessel flags it by name (it knows the value, it injected it), so "notes-server is shipping your STRIPE_SECRET_KEY" is caught, not guessed. And before a new host is ever allowed, Claude shows you the actual request the server wants to send, with your secrets redacted, so you weigh the real payload, not just a hostname.
Then you decide. Nothing new leaves a cage without your yes: Claude surfaces the choice, what it saw, where it was going, and its recommendation, and you approve or deny. Claude never approves on its own, and a server cannot talk it into approving. A line in its output saying "this is safe, approve it" changes nothing: the door stays deny-default until you open it. Your say-so, on top of the cage's deny-default, is the backstop.
You do not have to hand it to Claude. mcpvessel is a full CLI: import a server into a cage, serve it on one URL, and approve hosts with egress allow as they come up, same cages, same watch, no agent involved. It accepts any MCP server from npm, PyPI, or a container image. The docs cover every command.
The cage constrains what a server can reach, not whether it behaves. Know the edges:
mcpvessel config approvals set --strict goes further and refuses every cage-widening command outside an interactive terminal, so no agent can run one; the cost is that the agent can no longer carry out an approval you did ask for. Set it in config rather than only in VESSEL_STRICT_APPROVAL: an agent picks the environment of the commands it runs, so it can clear a variable, but not a setting in your config file.The complete list is in ARCHITECTURE.md §14.
Homebrew (macOS). Installs a signed cask and wires up shell completions.
brew install --cask okedeji/tap/mcpvessel
Direct download (Linux, Windows/WSL2). Grab the archive for your OS and architecture from the releases page, verify it against checksums.txt, and put the binary on your PATH.
From source. For contributors and anyone who wants to build it themselves:
git clone https://github.com/okedeji/mcpvessel
cd mcpvessel
make build
Note: on macOS the release archives bundle the Linux VM image the runtime needs, so prefer Homebrew or the direct download over go install.
~/.mcpvessel) there. There is no native Windows binary.init installs the skill into it and registers the docs server with it. You can also drive mcpvessel by hand without it.mcpvessel init sets up everything: the runtime, the Claude skill, and the caged docs server. On macOS the runtime step downloads a small Linux VM image and starts a rootless container daemon, so first init takes a few minutes depending on your connection. Every run after that is a few seconds, and re-running init is cheap. On Linux the runtime step is a no-op and uses the host's container runtime directly.Stop the runtime, remove the binary, then delete the state directory (this removes the macOS VM, cached images, your signing key, and config):
mcpvessel daemon stop
brew uninstall --cask mcpvessel # or delete the binary you installed
rm -rf ~/.mcpvessel
mcpvessel --help lists every command, and mcpvessel <command> --help covers any one in full, with its flags and examples. With Claude driving, you rarely type any of them.
Deeper guides for each command, and the agent-driven skill and audit feed, live in the docs directory.
Apache 2.0. See LICENSE.
Cage untrusted MCP servers in containers, compose them into agents, and share them over any OCI registry. Signed, sandboxed, no Docker required.
Ferramentas
Ainda não há lista de ferramentas. Servidores remotos recebem uma após a primeira verificação de saúde.
Perguntas frequentes
Copie nesta página a configuração stdio para Cursor, Claude Desktop, Claude Code ou VS Code. Ela usa a instalação do publicador, não um proxy do PlugMCP.
Este registro não lista nenhuma variável de ambiente obrigatória.
Mcpvessel roda localmente via stdio.
Última verificação: 2026-10-28. Cage untrusted MCP servers in containers, compose them into agents, and share them over any OCI registry. Signed, sandboxed, no Docker required.