mcp-ssh-fleet
by inhuman
README.md
# mcp-ssh-fleet
**English** | [Русский](README.ru.md)
[](https://github.com/inhuman/mcp-ssh-fleet/tags)
[](https://registry.modelcontextprotocol.io)
[](https://github.com/inhuman/mcp-ssh-fleet/pkgs/container/mcp-ssh-fleet)
[](https://github.com/inhuman/mcp-ssh-fleet/actions/workflows/docker-publish.yml)
[](https://go.dev/)
[](https://goreportcard.com/report/github.com/inhuman/mcp-ssh-fleet)
[](LICENSE)
[](https://github.com/inhuman/mcp-ssh-fleet/issues)
[](https://github.com/inhuman/mcp-ssh-fleet/commits/main)
MCP server that gives an agent two tools over SSH to a fleet of hosts outside
Kubernetes. The SSH key is a mounted secret (never enters model context); the
inventory is a fail-closed allowlist.
## Tools
- **`ssh_probe(tags, check)`** — curated read-only diagnostics on every host
carrying the given tags (AND semantics, like GitLab runner tags). `check` is
one of a built-in set: `uptime`, `disk`, `mem`, `failed`, `logs`. Arbitrary
commands cannot be passed. Class: `read-only`.
- **`ssh_exec(host, command)`** — one arbitrary non-interactive command on ONE
inventory host (by name or address; hosts outside the inventory are refused).
Class: `write-external`: the server executes the command, while access gating
(approval/RBAC) belongs to the MCP client.
Both: output size cap, per-host timeout, TOFU host-key check (fingerprint in
logs), output as one section per host.
## Configuration (env)
| Variable | Default | Meaning |
|---|---|---|
| `SSH_FLEET_TRANSPORT` | `http` | `http` (StreamableHTTP, endpoint `/mcp`) \| `sse` \| `stdio` |
| `SSH_FLEET_ADDR` | `:8080` | listen address for http/sse |
| `SSH_FLEET_AUTH_TOKEN` | — | optional `X-MCP-AUTH` token |
| `SSH_FLEET_INVENTORY_PATH` | `/etc/ssh-fleet/inventory.yaml` | path to the inventory (configmap) |
| `SSH_FLEET_KEY_PATH` | `/etc/ssh-fleet/id_ed25519` | path to the private key (secret) |
| `SSH_FLEET_OUTPUT_CAP_BYTES` | `8192` | output cap per section |
| `SSH_FLEET_CMD_TIMEOUT_SECONDS` | `20` | per-host timeout |
| `SSH_FLEET_PROBE_CONCURRENCY` | `8` | probe parallelism pool |
| `SSH_FLEET_PROBE_MAX_HOSTS` | `50` | safety cap on hosts per probe |
Inventory format — see `deploy/inventory.example.yaml`.
## Quickstart (docker, stdio)
```sh
docker run -i --rm \
-v /path/to/inventory.yaml:/etc/ssh-fleet/inventory.yaml:ro \
-v /path/to/id_ed25519:/etc/ssh-fleet/id_ed25519:ro \
-e SSH_FLEET_TRANSPORT=stdio \
ghcr.io/inhuman/mcp-ssh-fleet:latest
```
This is the shape MCP clients use when installing from the
[MCP Registry](https://registry.modelcontextprotocol.io)
(`io.github.inhuman/mcp-ssh-fleet`).
## Connecting an MCP client (http/sse)
By default the server speaks StreamableHTTP; register it as a regular HTTP MCP
server with endpoint `/mcp` (URL like `http://<host>:8080/mcp`). Both tools
become available to the client.
Client-side security recommendations:
- `ssh_probe` is `read-only` (curated checks only) and can be granted broadly.
- `ssh_exec` is arbitrary execution (`write-external`). Gate its access on the
client side (approval / RBAC / user allowlist) — the server merely executes a
command on an inventory host; it does not decide access policy.
## Development
```sh
make test # unit tests + e2e against a real in-process SSH server
make vet
make vulncheck
make build
make docker
```
Release: tag `vX.Y.Z` → GitHub Actions builds and publishes a multi-arch image
to `ghcr.io/inhuman/mcp-ssh-fleet`.
This server cannot be deployed
Maintenance
ActivityStale
ResponsivenessNo issues