Haru VPS MCP

A mini tunnel to a mini computer for ChatGPT.

Setup guide: Notion-style web note (中文 / English)

Haru VPS MCP is a small self-hosted MCP gateway for an isolated VPS workspace. It gives an MCP client a narrow filesystem/shell/file-transfer facade without making the host itself the workspace.

ChatGPT / MCP client
        |
 authenticated/private tunnel
        |
        v
 Haru MCP gateway
   127.0.0.1:8765
        |
   +------+------+
   |      |      |
   v      v      v
filesystem shell file-transfer
 backend   backend  child
 loopback  loopback loopback
   |      |      |
   +------v------+
 isolated workspace

Security boundary

Haru MCP exposes powerful workspace filesystem and shell tools, so treat the endpoint as privileged.

• The gateway refuses non-loopback bind addresses.

• Workspace backend URLs must be explicit loopback HTTP endpoints and cannot contain credentials.

• ChatGPT file ingress accepts only host-supplied file references, restricts downloads to approved OpenAI storage hosts over HTTPS, pins validated public DNS addresses before connecting, caps imports at 100 MiB, and writes only beneath the workspace root.

• Workspace file export accepts only workspace-relative regular files that resolve beneath the workspace root, caps exports at 100 MiB, and returns an MCP ResourceLink instead of creating a public download URL.

• Do not hand-craft file download URLs or treat raw client/sandbox paths as file references. The MCP host is responsible for supplying the file object declared through openai/fileParams.

• The backend itself gets no second public hostname; it stays behind the gateway on loopback.

• Optional public Host/Origin allowlists are request/transport hardening only. They are not authentication.

• deploy/Caddyfile.example fails closed with HTTP 403. Replace it only for a separately reviewed authenticated ingress design.

• For a private/on-prem/local MCP server used from ChatGPT, see docs/SECURE-TUNNEL.md and the current OpenAI Secure MCP Tunnel documentation instead of binding Haru to 0.0.0.0.

• Keep the delegated workspace disposable and separate from host configuration, credentials, home directories, and production data.

This public repository is a clean reference distribution, not a mirror of a private production host. It intentionally excludes private domains, machine identity, credentials, incident evidence, production deployment state, and owner-specific workspace contents.

Quick start: gateway

Python 3.10+ is required.

python -m venv .venv
. .venv/bin/activate
pip install -e '.[test]'
pytest
./deploy/verify.sh

deploy/haru-mcp.env.example is a non-secret environment template. The gateway reads its process environment directly; it does not auto-load a .env file.

For a local foreground run, copy/edit the template and explicitly export it into the shell before starting Haru:

cp deploy/haru-mcp.env.example .env
# edit .env as needed
set -a
. ./.env
set +a
haru-mcp

For systemd, install the reviewed values into the EnvironmentFile= used by your service unit instead.

By default, the gateway listens at 127.0.0.1:8765/mcp and delegates to:

http://127.0.0.1:8766/servers/filesystem/mcp
http://127.0.0.1:8766/servers/shell/mcp
http://127.0.0.1:8766/servers/file-ingress/mcp

Those endpoints are a separate workspace-backend composition. To build them from the selected upstream components plus Haru's bounded file-transfer child, follow docs/WORKSPACE-BACKENDS.md.

The public tool surface is deliberately small: gateway health, workspace directory listing/read/write/edit/move/stat, ChatGPT file import, workspace file export, and isolated shell execution delegated to the loopback backends. workspace_import_chatgpt_file is declared with openai/fileParams so the ChatGPT host can replace a current-conversation file with a short-lived file reference before the MCP call. workspace_export_file returns an MCP ResourceLink; a compatible client can read that resource and present the workspace file as a downloadable file.

Operator documentation

• docs/WORKSPACE-BACKENDS.md — build and operate the loopback filesystem/shell/file-transfer composition.

• docs/SECURE-TUNNEL.md — server-side Secure MCP Tunnel boundary, service supervision, fail-closed recovery, and real-client acceptance.

• docs/OPERATIONS.md — gateway service operations, layered health, upgrades/rollback, repository exact-head discipline, process hygiene, and secrets.

• deploy/haru-mcp.service.example — minimal hardened systemd starting point.

• deploy/Caddyfile.example — intentionally fail-closed public reverse-proxy example.

Upstream projects

The gateway imports the MCP Python SDK, AnyIO, and typing-extensions. The optional reference workspace composes mcp-proxy, the Model Context Protocol filesystem server, and shell-exec-mcp without vendoring their source.

Exact selected workspace versions/commits, the mcp==1.27.1 proxy-stack compatibility pin, and upstream license notes are recorded in THIRD-PARTY.md.

License

Haru VPS MCP code owned by this repository is available under the MIT License.

Third-party components keep their own licenses; see THIRD-PARTY.md.