lact-mcp — v0.5.0

MCP server (Python stdlib only, zero dependency) to control GPUs via LACT on Linux — so an AI agent can read stats and manage power / profiles / fan / clocks.

About

lact-mcp exposes the LACT daemon (socket + CLI) as MCP tools over stdio: monitoring first, safe writes second. Every write validates against hardware limits before touching the daemon, auto-confirms (lactd reverts unconfirmed changes after ~5 s), and tells how to restore stock settings. See CHANGELOG.md for the history.

Requirements

• Linux + LACT installed, lactd running (sudo systemctl enable --now lactd)
• User in the lact admin group (wheel by default): groups | grep wheel
• Python 3.10+ (stdlib only)

Run

python3 lact_mcp.py --test    # live self-check (needs lactd)
python3 lact_mcp.py           # MCP stdio server

Use with opencode (global MCP, no prompt)

Add to ~/.config/opencode/opencode.jsonc:

{
  "mcp": {
    "lact": {
      "type": "local",
      "command": ["python3", "/path/to/lact-mcp/lact_mcp.py"]
    }
  }
}

Then quit and relaunch opencode.

Tools (11)

Tool	What
list_gpus	List GPUs (index, PCI ID, model)
gpu_info	Static info: driver, VRAM, VBIOS, PCIe
gpu_stats	Live: clocks, voltage, power, temps, VRAM, fans, throttling
power	get / set power cap (set validates the configurable range first)
profiles	list / get / set / create (clone via from) / delete (refuses current) / import (PenguinBurner auto-uv JSON, apply to set current)
auto_switch	get / enable / disable auto profile switching
gpu_config_get	Full applied config (fan curve, clocks…) via lactd socket
daemon_query	Raw lactd socket call (device_stats, set_power_cap, set_fan_control…)
fan	get state / set static speed or curve (HW-minimum validated) / auto restore
clocks	get ranges+offsets / set min-max clocks, pstate offsets, boost, perf level / reset
voltage	get boost+VF curve / set per-point VF offsets, boost (NVIDIA) or offset/min/max (AMD) / reset

gpu_id accepts the short index ("1", or the number 1), the PCI address ("0000:01:00.0") or the full LACT ID — everywhere, including daemon_query args. Reads default to GPU 0; writes refuse a default when several GPUs exist (pass gpu_id explicitly, see list_gpus).

Examples:

// Fix fans at 50% on the RTX 3080, then back to automatic
{ "name": "fan", "arguments": { "gpu_id": "1", "action": "set", "mode": "static", "speed": 0.5 } }
{ "name": "fan", "arguments": { "gpu_id": "1", "action": "auto" } }
// +150 MHz core / +500 MHz VRAM on every pstate, then stock clocks
{ "name": "clocks", "arguments": { "gpu_id": "1", "action": "set", "gpu_offset": 150, "mem_offset": 500 } }
{ "name": "clocks", "arguments": { "gpu_id": "1", "action": "reset" } }
// Undervolt-style: lower VF point 0 to +25 MHz offset, then restore
{ "name": "voltage", "arguments": { "gpu_id": "1", "action": "set", "vf_points": { "0": 25 } } }
{ "name": "voltage", "arguments": { "gpu_id": "1", "action": "reset" } }

NVIDIA notes (proprietary driver)

• Fan control is usually unsupported (no hwmon) — fan get says so.
• Direct voltage is driver-locked — use voltage VF per-point offsets (the undervolt path) + power cap. set_clocks_value VF commands are AMD-only and silently ignored on NVIDIA.
• Daemon errors are returned in full (including the valid-command list).

Related

• LACT — the GPU controller this server drives. lact-mcp is an independent third-party MCP bridge (not affiliated); every write maps to a documented LACT socket/CLI call.

Safety

• power / fan / clocks set refuse values outside hardware limits first.
• Socket writes auto-confirm (no silent 5 s revert); every reply says how to restore.
• daemon_query GPU-config writes auto-confirm like the rest: every set_* command (incl. set_gpu_config) and batch_set_clocks_value (no silent 5 s revert).
• No secrets, no network: local socket + lact cli only.

Skipped (add when needed)

• snapshot tool — use daemon_query or lact cli snapshot directly.
• Remote TCP daemon — disabled by default in LACT (no auth); not exposed.