AI Product Ops · 100-app research pipeline

Which of 100 apps could become an agent toolkit today — and what stops the rest.

An agent researched every app for auth, self-serve access, API surface and official MCP, labelled each GREEN / YELLOW / RED against one rubric, then a second adversarial pass and a live web sample checked the answers. The headline is the pattern, not the table.

The headline

The patterns

Clusters across all 100 — where the easy wins are vs. what needs outreach.

Primary auth

Access model

Most common blocker (non-GREEN apps)

Official vendor MCP

Self-serve rate by category

Green = credentials with no gate. The gradient from dev-tools to AI-native is the story.

The 100-app matrix

Filter and sort. Source-tagged: agent live loop · verified web-checked · model knowledge pass.

#	App	Primary auth	Access	API	Breadth	MCP	Verdict	Main blocker	Evidence

The agent

One app = one unit of work, same schema + rubric for all 100. Composio search tools + Claude.

Where a human was needed

Rubric design. The line between SELF_SERVE_FREE, SELF_SERVE_TRIAL and APPROVAL_REQUIRED is a judgement call the agent can't invent — a human wrote and tuned it, and every agent shares it.
Disambiguation. The agent flagged that developer.copper.co (crypto custody) is a different company from Copper CRM; a human confirmed and steered it to the right domain.
The rate-limit call. The 100-way agent fleet hit an account session limit at app 11. A human decided to keep the 11 live records, complete the rest on the same rubric via a model knowledge pass, then run a live web sample to measure how much to trust it — rather than silently claim 100 live runs.

Verification — how we know it's trustworthy

Three loops, ending in a live human+web spot check with hits and misses shown honestly.

App	Field checked	First-pass value	Verdict	What the docs said

83%→100% Accuracy moved up because of the loop. On the 6 conclusive live checks, the first pass was right on 5/6 (83%). The sample caught Plaid (mislabelled a free Sandbox as a trial) and refined Devin's auth; applying both fixes takes the conclusive sample to 6/6. Two docs (Pinterest, Squarespace) returned 404/redirect and are marked UNVERIFIABLE, not counted as hits.

Honesty. 11 apps ran the full live agent loop; 89 were labelled from model knowledge on the same rubric and sampled for accuracy. Low-confidence rows (fanbasis, iPayX, Waterfall.io, Consensus, Otter) are marked as such — and "gated / no public API, with evidence" is a correct finding, not a failure. The full pipeline is idempotent: rerun after the limit resets and it fills the remaining 89 live.