diff --git a/README.md b/README.md
index c705dc48..32f5f858 100644
--- a/README.md
+++ b/README.md
@@ -168,6 +168,27 @@ This is useful for containerized or multi-node deployments where config is hoste
 
 > **Security best practice:** Never hardcode secrets in remote config files. Use environment variable references like `bot_token = "${DISCORD_BOT_TOKEN}"` and inject the actual values via local environment variables or Kubernetes Secrets. OpenAB expands `${VAR}` identically for both local and remote config.
 
+### SSH Sandbox (Local Deployments)
+
+For local deployments, you can run the agent inside an isolated VM or container using SSH as a transparent stdio transport — no changes to OpenAB are needed. SSH is a byte pipe over stdin/stdout, which is exactly how `AcpConnection` communicates with agents.
+
+```toml
+[agent]
+command = "ssh"
+args = [
+  "-T",                                     # no PTY — required, PTY corrupts JSON-RPC
+  "-o", "BatchMode=yes",                    # fail-fast, no interactive prompts
+  "-o", "ServerAliveInterval=30",           # keep-alive for long sessions
+  "-o", "ServerAliveCountMax=3",
+  "-o", "StrictHostKeyChecking=accept-new", # daemon has no terminal for prompts
+  "user@sandbox-host",
+  "claude", "--acp"
+]
+working_dir = "/tmp"                        # local cwd for the SSH process, not the remote agent's workdir
+```
+
+See [docs/ssh-sandbox.md](docs/ssh-sandbox.md) for setup details, MCP server access patterns, and known limitations.
+
 ## Configuration Reference
 
 > 📖 Full reference with all options, defaults, and Helm mapping: [docs/config-reference.md](docs/config-reference.md)
diff --git a/config.toml.example b/config.toml.example
index 277bdcf1..cc32c8ba 100644
--- a/config.toml.example
+++ b/config.toml.example
@@ -93,6 +93,22 @@ working_dir = "/home/agent"
 # working_dir = "/home/agent"
 # env = {}  # Auth via: kubectl exec -it <pod> -- cursor-agent login
 
+# SSH sandbox — run agent inside an isolated VM or container (local deployments)
+# SSH is a transparent byte pipe over stdio; no changes to ACP protocol needed.
+# See docs/ssh-sandbox.md for setup guide and known limitations.
+# [agent]
+# command = "ssh"
+# args = [
+#   "-T",                                     # no PTY — required, PTY corrupts JSON-RPC
+#   "-o", "BatchMode=yes",                    # fail-fast, no interactive prompts
+#   "-o", "ServerAliveInterval=30",           # keep-alive for long sessions
+#   "-o", "ServerAliveCountMax=3",
+#   "-o", "StrictHostKeyChecking=accept-new", # daemon has no terminal for prompts
+#   "user@sandbox-host",
+#   "claude", "--acp"
+# ]
+# working_dir = "/tmp"                        # local cwd for the SSH process, not the remote agent's workdir
+
 [pool]
 max_sessions = 10
 session_ttl_hours = 24
diff --git a/docs/ssh-sandbox.md b/docs/ssh-sandbox.md
new file mode 100644
index 00000000..adc32f17
--- /dev/null
+++ b/docs/ssh-sandbox.md
@@ -0,0 +1,124 @@
+# SSH Sandbox for Local Deployments
+
+OpenAB targets k3s on cloud, where Kubernetes NetworkPolicy and Pod isolation handle security. For local deployments (developer laptop, home server), the default config runs the agent with full host permissions:
+
+```toml
+[agent]
+command = "claude"
+args = ["--acp"]
+```
+
+The Claude subprocess inherits the host's full filesystem and network access. For a Discord bot accepting messages from arbitrary users, this is a meaningful attack surface.
+
+## SSH as a Zero-Code-Change Transport
+
+`AcpConnection::spawn()` treats the agent as a stdio JSON-RPC process. SSH is a transparent byte pipe over that same stdio — no changes to the ACP protocol, `SessionPool`, or `AcpConnection` internals are needed.
+
+```
+Current                              Proposed
+───────────────────────              ──────────────────────────────
+OpenAB                               OpenAB
+  │ spawn                              │ spawn
+  ▼                                    ▼
+claude (host permissions)            ssh -T user@sandbox
+  ├─ reads ~/.ssh ✗                    │ encrypted stdio pipe
+  ├─ reads ~/Documents ✗               ▼
+  └─ unrestricted network ✗          claude (inside sandbox)
+                                       ├─ restricted filesystem ✓
+                                       ├─ network allowlist ✓
+                                       └─ MCP via host proxy ✓
+```
+
+## Configuration
+
+```toml
+[agent]
+command = "ssh"
+args = [
+  "-T",                                     # no PTY — required (see below)
+  "-o", "BatchMode=yes",                    # fail-fast, no interactive prompts
+  "-o", "ServerAliveInterval=30",           # keep-alive for long sessions
+  "-o", "ServerAliveCountMax=3",
+  "-o", "StrictHostKeyChecking=accept-new", # TOFU on first connect; pre-populate known_hosts in production
+  "user@sandbox-host",
+  "claude", "--acp"
+]
+working_dir = "/tmp"
+```
+
+### Why `-T` Is Required
+
+| Flag | Behavior | JSON-RPC safe? |
+|------|----------|----------------|
+| `-T` | Clean byte pipe, stderr separated | Yes |
+| `-t` | Warns "PTY not allocated", stderr leaks into stdout | No — corrupts JSON stream |
+| `-tt` | Forced PTY + piped stdin → hangs indefinitely | No — deadlock |
+
+PTY inserts CR/LF conversion (`\n` → `\r\n`), merges stderr into stdout, and enables echo mode — all of which break JSON-RPC parsing. **`-T` is mandatory, not optional.**
+
+### `StrictHostKeyChecking=accept-new` and TOFU
+
+`accept-new` trusts the host key on first connection without prompting. This is safe for ephemeral local VMs but is trust-on-first-use (TOFU) semantics. For production environments, pre-populate `~/.ssh/known_hosts` manually:
+
+```bash
+ssh-keyscan sandbox-host >> ~/.ssh/known_hosts
+```
+
+Then switch to `StrictHostKeyChecking=yes` for stronger verification.
+
+### Why `BatchMode=yes`
+
+OpenAB runs as a daemon without a terminal. Interactive password prompts will hang the process. `BatchMode=yes` forces fail-fast behavior. SSH key-based auth must be configured beforehand.
+
+## Sandbox Options
+
+The SSH target is your choice — OpenAB does not care what is behind the SSH connection:
+
+| Environment | SSH target | Notes |
+|-------------|-----------|-------|
+| Mac (OrbStack) | `vm-name@orb` | Via `~/.orbstack/ssh/config` ProxyCommand |
+| Linux | `user@nspawn-container` | systemd-nspawn with SSH |
+| Remote machine | `user@10.0.0.5` | Any Linux server |
+| Docker | wrapper script using `docker exec` | Alternative to SSH |
+
+## MCP Server Access from Sandbox
+
+If MCP servers run on the host, the sandbox cannot reach them via `localhost` (which resolves to the sandbox's own loopback). Options:
+
+**Option A: Host DNS alias (OrbStack)**
+```
+claude (VM) ──http://host.internal:PORT──► MCP server (host)
+```
+
+**Option B: SSH port forwarding (universal)**
+```bash
+# add to ssh args:
+"-L", "8080:localhost:8080"
+```
+```
+claude (VM) ──http://localhost:8080──► [tunnel] ──► MCP server (host)
+```
+
+**Option C: Network bridge (Docker `--network host`)**
+```
+claude (container) ──http://localhost:PORT──► MCP server (host)
+```
+
+## Known Limitations
+
+### `kill_on_drop` does not reliably terminate remote processes
+
+Killing the local SSH client process leaves the remote subprocess running. The SSH server sends SIGHUP to the remote shell, but the agent may survive (especially with `nohup` or ControlMaster active).
+
+Mitigations:
+- Do **not** use SSH ControlMaster for agent connections. If your `~/.ssh/config` has `ControlMaster auto`, add `ControlMaster no` to the sandbox host entry.
+- Ensure the SSH server has `ClientAliveInterval` set to detect dead clients
+- Session pool TTL cleanup (`session_ttl_hours`) will eventually reclaim idle sessions
+
+### SSH connection startup latency
+
+Each `AcpConnection::spawn()` incurs an SSH handshake (~50–200 ms). This is negligible for long-lived sessions (default pool TTL = 24 h), but noticeable if sessions are frequently recycled.
+
+### SSH key auth is required
+
+OpenAB runs as a daemon without a terminal. Configure SSH key-based authentication on the sandbox host before using this transport.