First external Provider onboarding

[roy-tong]

First external Provider onboarding

本文件是 Issue #2 的正文源（docs/ISSUE-2.md）。push 时由
.github/workflows/issue-sync.yml 自动同步到
#2 —— 不要手改 Issue。

Goal

Onboard the first external capability provider onto AgentMeasure — a
remote/hosted/self-hosted MCP server or API that starts emitting canonical
observations instead of opaque call logs.

First deliverable: one fact about your server's usage you didn't know —
plus an honest statement of what cannot be known with today's lineage.

→ Lowest-commitment start — send a trace, get a measurement check:
20–100 anonymized trace/log rows (or a public export), mapped locally, with a
short report of what safely counts as attempts vs operations, where retries
may inflate usage, and what the telemetry cannot prove. No SDK, no
integration, no 7-day commitment; raw data stays local unless you explicitly
authorize sharing a sanitized sample (say whether your rows are production or
synthetic — separate intakes). Email tongroy18@gmail.com with subject
"measurement check" — three short fields (data, source + time window, the
decision it should inform).

→ Full 7-day audit — Provider Trial form
(five fields, ~2 minutes). Prefer email or this issue thread? Either works.
Full methodology:
product/AUDIT.md ·
Website.

Why this matters

Today every agent-facing tool keeps its own logs in its own shape, so nothing
is comparable. AgentMeasure's own baseline is
Pipeline Validation #001
(42 synthetic calls → 84 canonical observations, deterministic fixture,
deliberately 0 qualified usage). Measurement Report #1 is reserved for the
first external provider — your real traffic, your numbers, your story.

We are especially looking for providers who control the running server
process: remote MCP, hosted MCP, SaaS-backed MCP, or long-running
self-hosted deployments. (Local/stdio-only projects are welcome as SDK
implementers, but they don't see the call boundary.)

What onboarding looks like (≤30 min)

1. Pick a capability you expose (MCP server, API endpoint, CLI).
2. Wrap it with @agentmeasure/mcp (SDK v0.1.1, MIT) — install from the
   release tarball:
   npm install https://github.com/roy-tong/AgentMeasure/releases/download/v0.1.1/agentmeasure-mcp-0.1.1.tgz
   (npm registry publish pending scope/token), or clone and cd sdk && npm install.
   See sdk/examples/mcp-integration-v2.js (MCP SDK v2, primary) and
   mcp-integration.js (v1).
3. Emit canonical observations — unknown defaults are fine; honesty first,
   we never fabricate what can't be observed.
4. Done. The report is generated locally; nothing leaves your machine.

Two onboarding modes

• Private Alpha (default): instrument with the SDK, run 3–7 days on real
  traffic, and the report is shared only with you. Nothing leaves your machine.
• Public Alpha (opt-in): after seeing your private report, you may opt in
  to publishing your numbers as a Measurement Claim Label (multi-axis Evidence
  Profile — no ranking, no composite scores).

What you get

• One fact about your server's usage you didn't know — plus what remains
  unknowable and why.
• A say in where the standard's definitions break — your edge case becomes a
  spec change.
• Listed as an early implementer in the README (public only if you opt in).

Who this is for

Anyone who runs a remote/hosted/self-hosted MCP server or publishes an API for
agents, and wants real numbers about usage. If the SDK doesn't fit your stack
yet (Python? Go?), say so — that's exactly the feedback we need.

Reply here, or email the maintainer directly (see the private ops outreach
list). If you'd rather not install anything yourself, we can prepare the
integration PR for you — concierge onboarding is available.

[roy-tong]

Update: onboarding is now a 2-minute demo. ./examples/demo-e2e.sh runs the full chain locally (mock MCP server → 42 calls → 252 canonical observations → metrics report, 0 rejected). Wrap your own server the same way: sdk/examples/mcp-integration.js. Opening a PR or sharing output here gets you on the conformance list.