Open-source Revit MCP server: let Claude and other AI assistants read and edit BIM models, with human approval on every change.
Python aecai-agentsautodeskautomationbimclaudeifcmcpmcp-servermodel-context-protocolnavisworksrevitrevit-addinrevit-apirhino
ADR 0007: Hub Performance Posture
Status: Accepted Date: 2026-06-13
Context
As the AEC Model Bridge expands to orchestrate multiple workflows (such as W3 Model Health) and connect to diverse tools, there is a risk of creeping latency. We need a definitive performance posture that establishes the runtime language and execution budget to ensure the hub remains highly responsive without drifting into unnecessary full-system rewrites.
Decision
The hub stays Python. The Python ecosystem (MCP SDK, IfcOpenShell, specklepy, pyarrow/DuckDB) is our moat. Hub latency is typically dominated by host application responses and I/O, not interpreter overhead.
Rules for Hot Paths:
- Performance-critical operations must use native-backed libraries (e.g., IfcOpenShell's C++ core for geometry, pyarrow/DuckDB for the data lake,
orjsonfor large payloads). - A single module may be dropped to Rust/C++ as a Python extension (PyO3/nanobind) ONLY when a recorded profile demonstrates it dominates the workflow budget. Never attempt a whole-hub rewrite.
- Desktop switches (e.g., Revit, Navisworks) remain native by construction (C# in-process with the host).
Budgets
| Metric | Target Budget | Enforcement |
|---|---|---|
| Hub tool-dispatch overhead | < 10 ms | Perf smoke test (pytest -m perf) |
| W3 health audit on sample IFC | < 60 s | Continuous integration (upcoming) |
Consequences
- Agents are explicitly forbidden from initiating whole-hub language rewrites (e.g., to Rust or Go).
- We maintain a
perftest marker to assert our tool-dispatch budget programmatically. - Any new dependency processing large payloads must have a native core.