cad-mcp-cleanroom
Cleanroom acceptance harness that proves a CAD-style MCP command surface behaves safely and deterministically.
What it does and how it connects
cad-mcp-cleanroom is a cleanroom acceptance harness for CAD-style MCP command surfaces. It proves that a server's commands are deterministic, honestly declared read-only, and idempotent under replay, then emits a machine-readable PASS or REJECT report. The pitch is that if an agent will drive a CAD server unattended, this audit should run before letting it run anything.
The audited server implements a tiny protocol, execute(name, args) returning ok, data and error, plus state_hash(). Commands in a JSON manifest declare purity as read-only or mutating, and the harness checks determinism (run a read-only command N times and compare SHA-256 hashes of canonical output), purity (hash state before and after a read-only command), idempotence (apply a mutating command, hash, re-apply, hash), schema (output keys against declared types), and session drift (replay the manifest through two fresh sessions and compare outputs).
The README states its confidence plainly: high (85 percent) that replay-based purity and idempotence probing catches the mutation bugs agent loops actually hit, and medium (65 percent) that the bundled demo server resembles a real server's failure modes, both marked as estimates. The harness ships with a demo server; it audits a command surface, not a CAD design.
Alternatives
Compare side by side- ParamCAD-AgentBench
A tiered benchmark that scores agents building parametric CAD parts, rather than a harness that audits an MCP command surface.
- Text2CAD-GC
A geometric-continuity benchmark for text-to-CAD, rather than a cleanroom harness for MCP command surfaces.
- veriflow-rtl-benchmark
An open test kit that scores AI-generated Verilog against a golden model, rather than auditing CAD MCP command surfaces.
- llm-eval-ci-gate-rtl-verilog
A dual-tool CI gate for LLM-generated RTL, rather than a determinism and idempotence harness for CAD MCP servers.
- fail2bench
Turns agentic hardware design failures into new benchmark items, rather than auditing an MCP command surface before use.
Sources
Not verified
- Pricing: none stated on the repo.
- Real user voices: no real linked post read for this page; the repo has no public issues.
- No GitHub releases exist, so latest release fields are null.
Last verified November 6, 2026. Listings are informational. The evidence score measures public, verifiable evidence and is not an endorsement or a compatibility claim.