METHOD, read 10 Nov 2026
The directory lists and does not rank: each entry is read from its own source and dated
Download the directoryContext Weight Index method
The Data MCP Directory lists and does not rank. It records what each MCP server for a data tool publishes about itself and what can be observed in its source, field by field, and stops there. No page carries a total, an order of merit or a rating, and no page names a choice.
Every entry is read from the server's own repository, package registry page or vendor documentation. The version or commit read and the date of reading go with every entry. A server whose own source could not be read is not listed: on 10 Nov 2026 that rule admitted 30 servers.
The fourteen fields, in the order every entry shows them
| Field | What it records |
|---|---|
| Data tool | The product the server connects an agent to, such as Snowflake or Tableau. |
| Publisher | Who publishes the server, as its repository or documentation names them. |
| Kind | Official (the data tool's vendor), Official, Labs org (a vendor's Labs organization) or Community. |
| Repository or docs | The page the entry was read from. |
| Version or commit | The version, tag or commit read; for hosted services without versions, what the docs give. |
| Last release | The date of the latest release found, or none found on the read date. |
| Transport | Remote (a hosted endpoint over HTTP), Local (stdio on the user's machine), or both. |
| Auth | How the server authenticates, as documented: OAuth as the user, tokens, keys or service accounts. |
| Access | Read only, read and write, configurable, or not stated in the source read. |
| Tools listed | How many tool names the evidence records. |
| Tool definitions read at | Where the tool names and descriptions were read. |
| Context weight | Tokens in the published tool names and descriptions, from the Context Weight Index, or why it could not be measured. |
| Status | Whether it exists under that name, deprecations, renames and moves, as read. |
| Read | The date the entry was read. |
How kind, transport and access are labelled
- Official
- The data tool's own vendor publishes the server, in its own repository or documentation.
- Official, Labs org
- Published under a vendor's Labs organization on GitHub (AWS Labs, Snowflake Labs, Databricks Labs); the label says so because Labs projects can carry different support terms.
- Community
- Published by an individual or a group other than the vendor.
- Remote, Local, Both
- Remote means a hosted endpoint the client reaches over HTTP; local means the client starts the server as a process over stdio; both means the source documents both.
- Access
- Taken from the source: read only where the server documents only read tools or a read-only default; configurable where it documents switches, scopes or modes; read and write where write tools ship without a documented switch.
What counts as evidence
Every fact on the site carries a source URL and the date it was read. A publisher's statement about its own server is written as that publisher's claim. Where something was looked for and not found, the entry says none found on 10 Nov 2026, which records what the sources read showed on that date and is never written as a missing feature.
Facts are grouped on each entry by topic: publisher and releases; clients, auth and scoping; access and safety; what it reaches; lineage and impact; definitions, ownership and rules; freshness and upkeep. The topics are a filing system, and every entry uses the same one.
| Kind of fact | Facts | What it means |
|---|---|---|
| Vendor claim | 413 | The publisher's own documentation or README says it; written as the publisher's claim. |
| Observed | 56 | Read directly in the source code, the tool list, a package registry or the release page. |
| Third party | 1 | A source other than the publisher, such as a registry or another vendor. |
| None found | 52 | Looked for in the sources read and not found on 10 Nov 2026; never written as a missing feature. |
Why nothing is ordered by merit
Lists are filed alphabetically by data tool, then by server name, on every page and in every download. Kind and transport are labels a reader can filter on, not an order. Side-by-side pages put two servers' facts in the same rows and end without a conclusion.
The one sortable table on the site is the Context Weight Index, which is a measurement: it shows how many tokens each server's published tool definitions take, measured one way for all, and a reader can sort it by that number.
What the directory does not cover
The directory does not test servers against live accounts. It reads what each publisher documents and what the source shows, and says which.
MCP gateways, agent identity and agent governance platforms are outside its subject. So are data catalogs, lineage tools and context platforms as product categories: their servers are listed here, and the products themselves are not reviewed.
Sources
- The directory, CSV, read 10 Nov 2026
- The evidence, CSV, read 10 Nov 2026
Last reviewed 10 Nov 2026.