Evidence-governed screening for material decisions about physical assets — facility retrofits, equipment replacement, energy projects, vendor savings claims. Connect it to your AI assistant and it gains a governed screen: what the evidence currently supports, what could change the next action, and the justified posture — advance, condition, compare, measure, defer, or stop.
Remote MCP server (streamable HTTP, OAuth sign-in):
https://mcp.zircular.io/mcp
Listed in the official MCP registry as
io.zircular/zlab-stress-test.
New here? Start with What źlab does — a plain-language explanation of the problem it solves, what you get back, and a worked example.
Claude (web / desktop) — Settings → Connectors → Add custom connector → paste the URL. (Pro/Max)
Claude Code
claude mcp add --transport http zlab https://mcp.zircular.io/mcpCursor — add to ~/.cursor/mcp.json:
{ "mcpServers": { "zlab": { "url": "https://mcp.zircular.io/mcp" } } }VS Code — add to your mcp.json:
{ "servers": { "zlab": { "type": "http", "url": "https://mcp.zircular.io/mcp" } } }ChatGPT — enable developer mode, then Settings → Connectors → Create with the URL above (OAuth).
Full instructions: zircular.io/connect
For the plain-language "what does this actually do" answer, see What źlab does. In short:
Four tools: screen_operational_decision (submit; job-based, 3–10 min),
get_stress_test_status, fetch_stress_test_result, and a backward-compatible
alias. The result is an evidence-bounded governed read — never a categorical
yes/no, and financial magnitudes are withheld until site-level evidence supports
them. A refusal is a product answer: it names what is missing and how to lift
it, so your assistant can gather the named evidence and resubmit the same case to
raise the ceiling.
Host-facing documentation (also served in-band as MCP resources
zlab://docs/tool-contract and zlab://docs/document-workflows):
- Tool contract for hosts — the four tools, job lifecycle, payloads, refusal/error handling.
- Document workflows — when to invoke from documents/spreadsheets, what each posture means in the sheet, the evidence loop.
What leaves the machine, what is stored, and what is never logged: zircular.io/privacy. Every run requires explicit user authorization before public-data lookups (US Census, EIA, NOAA, EPA, and related public sources).
Built by źlab (Zircular LLC) — decision governance for physical assets. Questions: davidl@zircular.io
Note: this repository distributes the connector documentation. The analysis engine (64 governed motors) runs server-side and is not open source.