Skip to content

multi-instance.md

Latest commit

 

History

History
140 lines (110 loc) · 6.49 KB

multi-instance.md

File metadata and controls

140 lines (110 loc) · 6.49 KB
Raw

Multi-Maya-instance Deployment on a Single Workstation

Studios routinely run multiple Maya instances on one workstation — per-shot sessions for animation, a separate lookdev instance, a background batch renderer, and so on. This guide explains how dcc-mcp-maya behaves when N Maya processes share the same machine, and gives a drop-in userSetup.py recipe that keeps every instance discoverable through a single MCP gateway.

For cross-machine HA deployments, see dcc-mcp-core production deployment guide (core #330) — this page intentionally covers only the single-workstation case.

Invariants

  1. Every Maya instance must advertise a distinct dcc_pid so the diagnostics__* tools target the correct process. The default — os.getpid() — is almost always what you want; the only time you override it is when wrapping Maya with a launcher that proxies the real PID.
  2. Every instance reuses the same FileRegistry directory. In sidecar mode, the first Maya sidecar that notices the gateway is missing starts the standalone machine-wide gateway. All Maya processes then register as regular backends. Running two instances under different DCC_MCP_REGISTRY_DIR values makes them invisible to each other — only do this intentionally (e.g. per-user isolation).
  3. DCC_MCP_MINIMAL is per-process. One instance can run in minimal mode (just maya-scripting + maya-scene) while another runs full mode — the gateway reports each backend's active skill set independently.
  4. The gateway is independent of any one Maya process. If a Maya process exits, its sidecar exits and the gateway prunes that backend; the gateway itself stays up for other Maya, Blender, Photoshop, or future DCC sessions on the same workstation.

Launching N Instances

The bundled examples/multi-instance/userSetup.py leaves the instance port to core/OS assignment and registers with dcc_pid=os.getpid(). Drop it into your Maya scripts/ directory and every new Maya self-registers on next launch, without a probe-then-bind port race:

# examples/multi-instance/userSetup.py  (abridged)
import os


def _apply_multi_instance_env():
    os.environ.setdefault("DCC_MCP_GATEWAY_PORT", "9765")
    os.environ.setdefault("DCC_MCP_MAYA_DCC_PID", str(os.getpid()))

The full, commented source is in examples/multi-instance/userSetup.py.

Environment Variable Reference

Variable Scope Default Purpose
DCC_MCP_MAYA_PORT per-process 0 (OS-assigned) HTTP port for this Maya's MCP server
DCC_MCP_GATEWAY_PORT per-host unset (no gateway) Port the standalone gateway listens on
DCC_MCP_GATEWAY_NAME per-host dcc-mcp-gateway@<hostname> Label shown in admin and CLI gateway diagnostics
DCC_MCP_REGISTRY_DIR per-user platform default Shared registry every backend writes to
DCC_MCP_MINIMAL per-process 1 Minimal mode loads only maya-scripting + maya-scene
DCC_MCP_DEFAULT_TOOLS per-process unset Comma-separated list overriding minimal skill set
DCC_MCP_MAYA_HOT_RELOAD per-process 0 Watch bundled skills for edit-on-disk reload
DCC_MCP_MAYA_DCC_PID per-process os.getpid() Value advertised to gateway for diagnostics__* routing

The per-process variables can be set safely before Maya launches; the shared ones must match across every instance on the same user account, or they'll form separate registry islands.

Common Topology

┌────────────────────────────────────────────────────────────────┐
│  workstation-01                                                │
│                                                                │
│  Maya 2025.0 (anim)     port=<OS A>  dcc_pid=1234   minimal=1 │
│  Maya 2025.1 (lookdev)  port=<OS B>  dcc_pid=5678   minimal=0 │
│  Maya 2025.2 (batch)    port=<OS C>  dcc_pid=9012   minimal=1 │
│                              │                                 │
│                              ▼                                 │
│  shared FileRegistry ──► standalone gateway :9765             │
└────────────────────────────────┬───────────────────────────────┘
                                 │
                                 ▼
                          MCP clients

Every backend writes its (port, pid, dcc_version, skill_list, display_name) into the shared registry. The gateway keeps the routing table current and exposes a unified discovery surface that flags which backend owns which tool.

Troubleshooting

"I see only one instance from the gateway"

Check that all instances read the same DCC_MCP_REGISTRY_DIR. If one process launched under a different user account it will write into a different registry root. Use dcc_mcp_core.FileRegistry.list() to inspect the live entries.

"My instance never becomes the gateway"

In sidecar mode, Maya instances do not become the gateway. They ensure the standalone gateway is running, then register as backends. If you need a separate routing domain, set a unique DCC_MCP_GATEWAY_PORT and matching DCC_MCP_REGISTRY_DIR for that launcher.

"Stale entries after Maya crash"

The gateway's heartbeat prunes dead backends within ~30 s. To force a clean slate immediately:

# Remove the registry directory (recreated on next launch)
python -c "from dcc_mcp_core import get_config_dir; import shutil, os; \
           shutil.rmtree(os.path.join(get_config_dir(), 'registry'), ignore_errors=True)"

"One instance should run full mode, another minimal"

Set DCC_MCP_MINIMAL per-launcher, not globally. A Windows shortcut for the lookdev session might read:

set DCC_MCP_MINIMAL=0
"C:\Program Files\Autodesk\Maya2025\bin\maya.exe"

while the animation-pool launcher keeps the default (=1).

Related