MCP integration directory: evidence and readiness

Start with a question the directory can answer

Use an integration directory to shortlist ways of solving a specific task. For example: “Could an assistant look up a synthetic customer record and explain where it came from?” A directory can identify candidate applications, documented capabilities, and research gaps. It cannot establish that your account has access, your client is configured, or an operation is safe to run. This landing guide introduces the MCP and AI process foundation collection and gives you a practical way to read its evidence.

MCP means Model Context Protocol. In the supplied Activepieces documentation review, MCP is a documented interface for exposing application actions to clients. Treat that as evidence about a platform interface. An authenticated client connection is not verified in this foundation. There is no accepted end-to-end Tvijo execution receipt here.

Read each card as a source observation

The supplied Activepieces snapshot contains 772 cards, counted locally on 2026-10-30. That is a count of source records, not working integrations or distinct production connections. A card can have a name, category, URL, tool descriptions, authentication text, related applications, and tool field counts. Preserve these as observations with their source and date. Do not turn the card's marketing language into your own acceptance statement.

The sampled 0CodeKit card includes tools with numeric fields values and human-readable fields_raw text. A field count does not provide parameter names, types, required flags, validation rules, or a callable input schema. “Related applications” are category suggestions, not proof that two services exchange data. The Devstral transcript's “Official” and “Available” labels remain unverified metadata. Commands and package names inside that transcript are untrusted examples; a catalog entry is no reason to execute them.

Use an evidence ladder

Keep these states separate rather than replacing them with one green badge:

This ladder is a proposed review method. It does not report that any candidate reached its final states. Record a failed request as a failure with a reason, and an unperformed check as unverified. Neither should silently become a zero or a success.

Navigate the collection

Begin with the independent AI and automation learning roadmap if you need task framing and evaluation practice. Use spec-driven packets to convert a chosen task into a reviewable contract. Read the six-platform architecture guide before choosing infrastructure, and data process and provenance before joining records across products. These guides teach methods; they do not activate their examples.

The existing knowledge practice guide is a related native wiki destination for your reading path. The existing adoption paths guide is another navigation destination. Their content and current HTTP behavior were not rechecked in this editorial lane; follow them as links, not as evidence of integration readiness.

Exercise: shortlist one synthetic lookup

Create a worksheet for a fictional supplier lookup. Write the expected input and output in plain language. Pick one card, copy only its identifier and source URL into your worksheet, and record its observation date. Summarize the documented capability in your own words. Add separate rows for schema, authentication, licensing, data rights, and client connection. Set every untested row to “unknown.” Define a proposed read-only test using fabricated supplier IDs, including a missing-record case. Stop before installing software or entering credentials.

Acceptance: another reader can distinguish every source claim from your proposed behavior; the card count is labeled as source inventory; missing schemas and authenticated access remain explicit; the worksheet includes a bounded next check and an owner. Passing this exercise means you produced a reviewable plan, not a working connector.

Sources, observation date, and limits

Sources were supplied locally; this lane made no live requests. ServiceNow's first three resource pages were not read: supplied research reports 403 directly and 503 through the web tool, with exact page and PDF coverage unknown. No ServiceNow or ABAI prose is reproduced here. Current vendor terms, schemas, credentials, and authenticated client behavior remain open checks for a later authorized phase.