MCP Server¶
The MCP server exposes pacsys device read/write as tools for AI agents (Claude Code, etc.) via the Model Context Protocol.
Overview¶
Three tools are exposed:
| Tool | Description |
|---|---|
read_device(drf) |
Read a device value |
write_device(drf, value) |
Write with policy enforcement |
device_info(name) |
Look up device metadata from DevDB |
Write safety comes from two layers:
- Claude Code's tool permission prompt — human-in-the-loop approval for each write call
- Server-side policy chain —
DeviceAccessPolicy→ValueRangePolicy
Without a policy config, all writes are denied. Reads are always allowed — the TOML config only expresses write policies, and create_server accepts no others. This is the same policy system used by Supervised Mode; for read restrictions, run the supervised proxy instead.
Writes require Kerberos credentials. The server refuses to start if write devices are configured but no Kerberos ticket is available.
Quick Start¶
Read-only (no config needed)¶
Where scripts/.mcp_prod.json contains:
{
"mcpServers": {
"pacsys": {
"type": "stdio",
"command": "python",
"args": ["-m", "pacsys.mcp"]
}
}
}
Or register it with the CLI:
With write access¶
Where scripts/.mcp_prod.json points to a TOML config:
{
"mcpServers": {
"pacsys": {
"type": "stdio",
"command": "python",
"args": ["-m", "pacsys.mcp", "--config", "scripts/pacsys-mcp.toml"],
"env": {}
}
}
}
Configuration¶
TOML config file¶
[server]
transport = "stdio" # "stdio" or "sse"
# port = 8000 # only used with transport = "sse"
# role = "testing" # DPM role for access control
# audit_log = "audit.jsonl"
[policies]
# Devices allowed for writing (glob patterns).
# Without this list, ALL writes are denied.
write_devices = ["Z:ACLTST", "Z:CUBE_Z"]
# Raw writes to range-limited devices are denied unless explicitly exempted.
allow_raw = ["B:HS*"]
# Value range limits per device pattern.
[policies.value_ranges]
"Z:ACLTST" = [0.0, 100.0]
A sample config is provided at scripts/pacsys-mcp.toml.
Configuration is validated strictly at startup. Unknown keys, malformed policy
tables, invalid bounds, and transport-specific options such as port with
stdio stop the server instead of silently weakening a policy.
When audit_log is configured, every allowed or denied write attempt is
recorded as JSON Lines with its requested DRF, value, decision, and reason. The
file is opened during startup so an invalid audit destination prevents the
server from accepting clients, and a write whose approval cannot be recorded
is blocked rather than executed unlogged.
CLI flags¶
CLI flags override values from the config file:
| Flag | Description |
|---|---|
--config PATH |
TOML config file |
--transport |
stdio (default) or sse |
--port |
Port for SSE transport |
--role |
DPM role for access control |
--debug |
Enable debug logging |
Environment variables¶
The server respects standard pacsys environment variables:
| Variable | Description |
|---|---|
PACSYS_DPM_HOST |
DPM proxy hostname (default: acsys-proxy.fnal.gov) |
PACSYS_DPM_PORT |
DPM proxy port (default: 6802) |
These can be set in the MCP config JSON:
{
"mcpServers": {
"pacsys": {
"type": "stdio",
"command": "python",
"args": ["-m", "pacsys.mcp"],
"env": {
"PACSYS_DPM_HOST": "my-proxy.fnal.gov"
}
}
}
}
Tools¶
read_device¶
Read a device value. Accepts any valid DRF string.
Parameters:
| Name | Type | Description |
|---|---|---|
drf |
str |
DRF string (e.g. M:OUTTMP, M:OUTTMP.SETTING, M:OUTTMP[0:9]) |
Returns:
{
"ok": true,
"name": "M:OUTTMP",
"drf": "M:OUTTMP",
"value": 72.5,
"units": "deg F",
"timestamp": "2026-03-03T14:00:00+00:00",
"cycle": 0
}
On error:
{
"ok": false,
"name": "M:BADDEV",
"drf": "M:BADDEV",
"value": null,
"facility_code": 16,
"error_code": -42,
"error": "DIO_NO_SUCH - device not found"
}
Any non-zero status adds facility_code and error_code. A warning (error_code > 0) with usable data stays "ok": true and carries its text in message; failures carry error (the status message, or "Read failed (facility=F, error=E)" when there is none).
write_device¶
Write a value to a device. Requires policy approval.
Parameters:
| Name | Type | Description |
|---|---|---|
drf |
str |
Device DRF string |
value |
float, str, or list |
Value to write |
Returns:
On denial:
{"ok": false, "drf": "Z:ACLTST.SETTING@N", "error": "Value 200.0 for Z:ACLTST outside range [0.0, 100.0]"}
A backend write failure also includes the status codes; error falls back to "Write failed (facility=F, error=E)" when the server sends no message:
{"ok": false, "drf": "Z:ACLTST.SETTING@N", "facility_code": 17, "error_code": -44, "error": "Write failed (facility=17, error=-44)"}
device_info¶
Look up device metadata from the device database.
Parameters:
| Name | Type | Description |
|---|---|---|
name |
str |
Device name (e.g. M:OUTTMP) |
Returns:
{
"ok": true,
"name": "M:OUTTMP",
"description": "Outside temperature",
"device_index": 12345,
"reading": {"units": "deg F", "common_units": "deg F", "min": -40.0, "max": 140.0},
"setting": {"units": "deg F", "common_units": "deg F", "min": 0.0, "max": 100.0},
"control_commands": [{"value": 0, "short_name": "OFF", "long_name": "Turn off"}]
}
DevDB must be available (requires grpcio). If unavailable, returns {"ok": false, "error": "DevDB client unavailable"}.
See Also¶
- Supervised Mode — the policy system used by the MCP server
- Writing Guide — write operations and authentication
- DRF Format — DRF string syntax for device addressing