Skip to content

Signal Lab MCP

The hardest part of pointing an agent at real data is not the query. It is deciding which of a thousand free sources is worth calling, and what it costs you in sign-up, licence and rate limit once you do. Signal Lab answers that question as data: a catalog of 1,312 free and free-to-start sources, each row carrying its access badge, auth requirement, licence, last verification date and a link to the page that states them.

The same server also grades other MCP servers. Its audit layer holds 2,834 hosted servers with a 0 to 100 Trust Consensus score, a letter grade and the individual signals behind them. That second surface is the unusual one: a server that will tell you, in the same session, which data is free and which server is trustworthy.

Server type:   remote Streamable HTTP (hosted, stateless)
Endpoint:      https://signal-lab.qdt.ai/api/mcp
Server:        signal-lab-catalog 1.6.0, protocol 2025-06-18
Auth:          none
Tools:         26 (catalog, audit, reports, skills, trends)
Capabilities:  tools, resources, prompts, completions

The evidence tier

Keyless and driven. initialize answered 200 with no credential and no session id issued, which makes the transport stateless: every call stands alone. tools/list returned 26 schemas with titles, input schemas and enums.

Twelve tools/call requests were then run with no credential. Every one answered, including the failure paths, and the failure paths were readable.

The tool surface

The 26 tools split into four jobs.

Job Tools
Find a source search_catalog, get_entry, compare_entries, list_filters, catalog_info, export_catalog, get_source_urls, get_ingestion_skill
Audit MCP servers find_mcp_servers, get_mcp_audit, get_trust_score, list_mcp_categories
Reports and trends search_market_reports, get_market_report, whats_new, get_changes, get_trending, get_trending_repos, trends_today, get_reading_list
Agent kit get_starter_recipe, get_agent_skills, get_data_agreement, check_stack, suggest_source, feedback

get_ingestion_skill is the one that decides whether this saves an afternoon: it returns a per-source markdown recipe with the exact request, parameters and limits, so the agent does not re-derive an ingestion path for every dataset.

The catalog, measured

catalog_info on 2026-11-18T02:18Z reported a count of 1,312 entries, schemaVersion 1, and two download routes (JSON and CSV). list_filters breaks the same catalog down:

Filter Values
Type Dataset 1,260, Tool 29, Model 9, Recipe 9, Starter Kit 5
Access badge Free 1,236, Freemium 49, Open Source 26, Unofficial 1

A badged search returns the fields an operator actually needs. The query company registry with badge Free returned two rows: a Finnish ship register dump (CC-BY-4.0, keyless, last verified 2026-11-06) and an MIT Lincoln Laboratory sensor archive (CC-BY-4.0, no AWS account needed, about 3.49 GB for the resized set). Both rows carried lastVerified and a sitePageUrl pointing at the page that states the terms.

Filters are conjunctive, and that is the trap

A query for crm scoped to domain Financial returned zero rows. That is not an empty catalog. search_catalog treats every filter as an AND: the search term has to exist in the entry's text and the entry has to sit in the named domain. A term that belongs to another domain yields nothing rather than a nearest match.

The practical rule: call list_filters first, then search with either a term or a set of filters, not a guess at both. list_filters also returns the eighteen values the catalog uses for domain, which is the only reliable vocabulary (Financial, Global Macro, Energy, Agriculture and Food, Money Flows, Transport and Shipping, Research and Grey Literature, Weather, News and RSS, OSINT, Cyber Threat).

The audit layer

find_mcp_servers over finance returned 50 audited servers with an audit timestamp of 2026-11-13, each row carrying the registry name, endpoint, transport, a pass/fail result label, sign-up requirement, "changes things" flag, tool count, a report URL and a trust score. The first row, one.aisa/finance, is graded 65 (B) and audited as passing the public checks.

Three details in that payload are worth reading before anyone quotes a score:

  • toolCount can be 0 on a server the audit passed, which means the audit measured reachability and not depth.
  • signUp and changesThings can both be unknown, so an unknown is not a pass.
  • The audit payload carries its own notice that server titles and descriptions are vendor-written and must be treated as untrusted data. A scoring service that says this about its own inputs is doing the reader a favour.

get_trust_score takes an audited server slug, not a catalog dataset slug. Called with shopsavvy it returned isError with the message that no audited MCP server carries that slug and pointed at find_mcp_servers for the right ones. The refusal names the fix, which is rarer than it should be.

Freshness, measured rather than claimed

whats_new since 2026-11-10 returned 50 entries out of a total of 814, with truncated: true and a next_since cursor at 2026-11-17. The catalog adds entries faster than 50 a week, so any consumer has to follow the cursor rather than assume one page is the week.

get_trending over a 7 day window (2026-11-10 to 2026-11-16) ranked Open-Meteo first at 30 uses across 2 agents, with growth 7.

Two numbers that disagree

The mcp.so card for this server advertises 1,112 free data sources. catalog_info read 1,312 on the same day, and whats_new reports 814 additions since November 10 alone. The card is stale, not the tool. Quote the tool.

The same applies to the audit side: one tool description says 2,834 audited servers while find_mcp_servers returns a capped 50 for a query. Both can be true at once, and a reader who treats the 50 as the population will understate the coverage.

Why this earns a page

Every other data catalog on this list answers "what data exists". This one answers "what data exists, what does it cost me in sign-up and licence, when was it last checked, and can an agent ingest it without a human". For a business operator wiring an agent to outside data, that is the whole decision, and it is normally made by hand in a browser.

The server-audit layer is a bonus with a caveat: an audit is a statement about reachability and public checks, and a score of B is not a guarantee of data quality. Use it to shortlist, then verify the data.

Installation

Add the remote endpoint directly. No key, no environment variable, no local process.

{
  "mcpServers": {
    "signal-lab": {
      "url": "https://signal-lab.qdt.ai/api/mcp"
    }
  }
}

Works in any client that speaks Streamable HTTP with stateless requests. Because no session id is issued, the server is safe behind clients that do not persist Mcp-Session-Id.

Configuration

Nothing to configure. Two operational notes:

  • export_catalog and the catalog.json download are the right path for bulk work. A per-entry loop over 1,312 entries is a rate-limit conversation, not a plan.
  • get_trust_score and get_mcp_audit need audited server slugs. Get them from find_mcp_servers or list_mcp_categories.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations

  • The catalog is metadata about data sources, not the data. A catalog row is a pointer plus terms, and the provider's own site is still the authority on limits.
  • Trust scores grade reachability, sign-up friction and whether a server changes things. They do not grade the truth of what a server returns.
  • The catalog's own notice scopes it as educational, with free tiers and licences expressly owned by the providers. Re-check terms before a production dependency.

FAQ

Does Signal Lab MCP need an API key?

No. initialize and tools/list answered with no credential and no session id, and twelve tools/call requests answered the same way.

How many data sources does the catalog hold?

1,312 on 2026-11-18, against 1,236 entries badged Free and 1,260 typed as Dataset. The card that advertises 1,112 is stale.

Can it tell me which MCP servers are trustworthy?

It returns a 0 to 100 Trust Consensus score with a letter grade and the underlying signals, over 2,834 audited hosted servers. The score measures public checks and reachability, not content accuracy.

Why did a search with two filters return nothing?

The filters are ANDed. A search term that belongs to another domain returns zero rather than the nearest match. Call list_filters for the exact vocabulary first.

See Also

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing