How to Use CapaLens: for People and for AI

Scope: this article retains its October 2026 viewpoint and release context. Design goals, plans, examples and delivered capabilities are distinct; availability depends on the release in use.

This part is the practical guide. There are two paths: people search on the website, and AI queries over MCP. Both read the same Intelligent Access (IA) catalog, so they see the same facts.

1. Three steps: describe the task → compare conditions → review, then connect

Step one: describe the task. Start from the work you want done, not from a tool name. The search box on the home page asks for "your business task or a tool," and shows real examples underneath: research public company information, find authorized customer records, review customer service tickets, check order status, review contract signature status, review payment transactions, review project progress, search authorized team knowledge.

Include your conditions and the results get sharper — for example, "self-hosted web analytics without cookies."

Step two: compare conditions. Compare capabilities, environment, authorization, pricing sources and known limits. The current comparison contract accepts two to four distinct record IDs. Filters and fields must be taken from the release in use; an unknown condition is not a claim of support.

Step three: review, then connect. Go back to the official documentation for the latest conditions, review the configuration, and connect from your own client directly to the vendor. CapaLens organizes documentation and evidence; it does not operate vendor services or hold vendor keys.

2. Reading a result on the website

The following result-card structure is a design direction, not a list of features all delivered in v0.4.8 or a promise attached to a release date:

• Function-oriented titles could show a task above the product and interface metadata.
• Condition states should distinguish met, unmet and unknown; missing evidence must not become an assumed match.
• Evidence labels need specific methods and scope. Unreviewed listings and documentation records must not be called business-tested; human review is not automatically an end-to-end call test.
• Freshness indicators, expiry and version-change review are proposed behavior. An automatic downgrade requires implementation and evidence.
• Ranking explanations should cite recorded mappings and fields; publishing this article does not establish that every ranking rule is already public.
• Labels such as "safe," "guaranteed" or a blended star score must not replace the actual evidence.

When a response provides an index digest or snapshot date, retain its scope so a person and AI can compare the same data. This article does not claim that every page footer already displays those fields.

A detail can show source links, recorded dates, business mappings, known limits and IA guidance. Dates and pricing fields may be unknown. Formal pricing projections and supplementary source claims have different evidence scopes; neither implies that every product has four known price layers.

3. For AI: querying over MCP

CapaLens runs a read-only MCP server:

• Endpoint: https://capalens.com/mcp
• Transport: Streamable HTTP
• Auth: none required for the public catalog

The following is an HTTP MCP configuration example for Claude Code. Other clients may use different configuration structures; check their current official documentation first:

{
  "mcpServers": {
    "capalens": {
      "type": "http",
      "url": "https://capalens.com/mcp"
    }
  }
}

For Claude Code, a command example to review, not run by this page:

claude mcp add --transport http capalens https://capalens.com/mcp

For apps such as Claude or ChatGPT, first check whether the current client, plan, workspace policy and administrator permissions support custom remote MCP, then follow the official setup flow. Support is not identical across desktop and web clients. Adding CapaLens may establish a connection to its MCP endpoint; it does not automatically install or connect the other vendor services listed in its catalog. Your AI gains a place to check evidence.

The following is a target interaction, including condition states and follow-up information designed for later releases. It does not imply that v0.4.8 returns all these fields:

You say: "Connect me a tool that can look up support tickets — official source, OAuth." In the target interaction, the AI calls search_software and presents candidates, condition states, evidence and unknowns using fields actually delivered. Three-state conditions and suggested follow-up questions depend on the relevant release output. For an unknown condition the AI should come back and ask you ("does the data have to stay in-country?") rather than guess. Once you pick, you confirm the addition in your own client and your credentials go straight to the vendor. Later calls never touch CapaLens.

The tools

The preparation snapshot retains 15 read-only tools: seven formal-catalog tools, one work-path assessment tool and seven source-candidate tools. This is the local contract used for the article, not a new observation of every production call:

Two practical tips:

1. Start with a bounded search, then read a chosen ID. Use the documented compact search result and a small limit; do not invent arbitrary fields values. For get_software, the optional field selector is restricted to pricingV2, selectionGuide and businessNeeds; omitting it requests the bounded full record.
2. Keep data scopes separate. The formal catalog has 246 software entries and 363 interface records; 11,382 Registry listings are counted separately and remain unreviewed/untested. Six supplementary information records are local preparation, not a claim of a larger live unique-product count. Official Registry provenance does not make every server vendor-official. The homepage already queries the catalog and Registry in separate panels; canonical cross-source identity reconciliation remains distinct.

Read the scope and unknowns that the actual response provides. CapaLens supplies directory information and evidence, not a combined execution plan, safety guarantee or vendor call proxy. "Unknown" means no evidence for that field, not proof of non-support.

4. Taking the result with you

Copy for AI is a design direction subject to implementation and release verification. A proposed compact text block would include an interpreted query, candidate IDs, condition states, source links and scope. The following is an illustrative layout, not a real checked Notion record or proof that the button is delivered:

[Illustrative CapaLens access-card format; version and date are placeholders]
Notion MCP (official, remote MCP, OAuth required)
Capabilities: search workspace; read and update pages; create pages and databases
Conditions: official source ✓ | read-only access ✗ (includes write) | data in-country ? (no evidence yet)
Evidence: developers.notion.com official docs (documentation review, not call-tested)
Scope: CapaLens provides evidence only; you confirm connection and authorization in your client.
Re-check: CapaLens MCP get_software id=notion

The last line tells the receiving AI how to come back to CapaLens and verify — so what you forward is not a conclusion but a verifiable lead.

Access cards and "use it in my AI" are proposed features. One possible design is a stateless link that shows conditions and evidence before the user reviews a client-specific addition. This is not an available one-click workflow established by the article.

The proposed flow is to choose a client, review the source-declared server address and configuration, and confirm any addition in that client. Install links, commands and fallback configuration depend on the client and release; none is executed by this article. API or CLI guidance would remain reviewable documentation rather than a request for CapaLens to run the tool.

The red lines do not move: no installing on your behalf, no credential collection, no call relaying. Every addition is confirmed in your own client.

5. Common questions

I cannot find a match. Text vocabulary, recorded mappings and source coverage may limit results; no result is not proof of non-support. Try a product name, an English keyword or a documented business-task ID such as review-customer-tickets. Richer need routing is a development direction, not a release-date guarantee. Publishing this article does not enable query logging, feedback collection or new telemetry; those require their actual implementation and privacy scope.

What does "unknown" mean? We have no evidence — not that the feature is unsupported. Relax that condition and search again, check the official docs, and tell us through the correction channel.

How often is information updated? Use a recorded date only with its stated basis. Some dates are absent; a source research date is not a vendor business test. Continuous monitoring and automatic rechecks are design directions, not established behavior of this article. Check current vendor documentation, especially permissions and pricing.

Does CapaLens cost money? Public queries are subject to the site's limits. Subscriptions for continuous checks, alerts or team records are business hypotheses, not billing introduced by this article. Payment must not buy a verification result; sponsorship, if offered, needs clear labeling and separation from evidence and natural results.

---

This guide uses a preparation snapshot of v0.4.8 contracts. It does not replace the current tool schema or a fresh production check. Features described as design or planned are not made available by this article.