Nikita Shamgunov, a Neon co-founder who is now a Databricks vice president of engineering working on its Lakebase Postgres, posted “Quality wins” on X at 7:36pm UTC on 20 November, which was 8:36am on 21 November in New Zealand. The link in the post opens Manufact’s MCP quality leaderboard, filtered to data and analytics, where Neon’s MCP server sits first at 93 out of 100.
The post is two words and a link, with no comment from Shamgunov on what the ranking means. Manufact, which runs the board, is a startup that sells hosting, testing and analytics for MCP servers. The report behind Neon’s number shows where the 93 comes from and what it leaves out.
- Who posted it, and why MCP matters to Neon
- What Manufact measures
- Neon’s report card
- The data and analytics board
- What the score doesn’t measure
- What the server does and how to narrow it
- Sources
Who posted it, and why MCP matters to Neon
Neon has been part of Databricks since May 2025, when Databricks announced it had agreed to buy the serverless Postgres company. A month later Databricks launched Lakebase, which it says is built on Neon technology. Neon’s site says agents can operate it through MCP, a command line tool and an SDK, and a Databricks summary of a talk Shamgunov gave after the acquisition says 80% of the databases created on Neon come from AI agents, not people. That’s why the quality of Neon’s MCP server, the interface an assistant uses to call a service’s tools, matters to the company.
What Manufact measures
Manufact scores an MCP server by connecting to it and reading what it advertises. Its methodology post, published on 23 October by co-founder Pietro Zullo, lists five categories at 20% each: tool documentation, tool definitions, protocol compatibility, server metadata and discoverability. A server that ships MCP Apps gets a sixth category, and each of the six then counts for about 16.7%.
Manufact says the review reads tool metadata, app resources and public web pages without executing a server’s tools. Discoverability has six equal checks. Four look for a listing in Claude’s connector directory, the ChatGPT apps directory, the Cursor marketplace and Muse, and two look for an mcp.json manifest on the company’s domain and setup instructions served at the MCP address. A Verified Claude listing earns full credit on its check and a Community listing earns half.
| Category | Weight | What it looks at | Neon’s score |
|---|---|---|---|
| Tool documentation | 20% | Titles, descriptions, inputs, output schemas and behaviour annotations | 80 |
| Tool definitions | 20% | The structure of the tools the server exposes | 100 |
| Protocol compatibility | 20% | Whether the server negotiates and exposes MCP capabilities as expected | 100 |
| Server metadata | 20% | Information that helps a client and a developer understand the server | 100 |
| Discoverability | 20% | Store listings and public setup information | 83 |
- MCP server
- A service that lists tools an AI assistant can call, using the Model Context Protocol.
- Output schema
- A machine-readable description of what a tool returns.
- Annotations
- Four hints on a tool: whether it only reads, can destroy data, can be repeated safely or reaches outside systems.
- Lakebase
- Databricks’ Postgres service, which Databricks says is built on Neon’s technology.
- Discoverability
- Manufact’s name for whether people can find and install a server outside its own site.
93
Neon’s overall score, grade A
6th
overall, of 1,674 curated servers
141
tools in the documentation review
68.3
average score across 72 data and analytics servers
Neon’s report card
Neon’s report, last reviewed at 06:16 UTC on 20 November1, gives it 100 for server metadata, 100 for protocol and 100 for tool definitions. Discoverability scores 83: five of its six checks passed and one is marked unverified, which earns no points until Manufact confirms it. Tool documentation scores 80. The five numbers average 92.6, which rounds to the 93 on the board.

The documentation score covers all 141 tools Manufact discovered2. The lowest-rated is list_branch_computes at 65. Eight tool reviews on the report show titles scoring between 90 and 99 and every behaviour annotation, the four hints that say whether a tool only reads, can destroy data, can be repeated safely or reaches outside systems, at 100. Their output-schema scores, which grade the machine-readable description of what a tool returns, run from 19 to 63, and recover_project is lowest at 19.
Separately, Manufact’s publishing checklist gives Neon 100 out of 100 as of 20 November and says the scan meets both ChatGPT’s and Claude’s publishing checks. Manufact says passing doesn’t guarantee marketplace approval. Neon’s documentation calls the server an official Claude connector, and Manufact lists it as a partner in Claude’s directory.
The data and analytics board
In Manufact’s data and analytics category, Neon’s 93 is followed by Polygraph at 88, HYPD AI for Google Analytics at 87, ClickHouse Cloud’s MCP server at 86 and Crash Round Lab at 85. Manufact’s statistics put the category average at 68.3 across 72 servers, and the overall average at 70.2 across 1,195 scored servers.

| Category rank | Server | Score | Overall rank | Last reviewed |
|---|---|---|---|---|
| 1 | Neon | 93 | 6 | 20 Nov |
| 2 | Polygraph | 88 | 40 | 12 Nov |
| 3 | HYPD AI for Google Analytics | 87 | 57 | 15 Nov |
| 4 | ClickHouse Cloud MCP Server | 86 | 68 | 17 Oct |
| 5 | Crash Round Lab | 85 | 91 | 11 Nov |
Neon’s score has changed. An earlier snapshot of Manufact’s catalogue, from when it listed about 1,000 curated servers, shows Neon’s server at 83 with a B grade, third in the category behind ClickHouse at 86 and Axiom at 84. ClickHouse’s entry still reads 86. Neon’s MCP documentation was last updated on 17 November and Manufact reviewed the server again on 20 November. Manufact says older reports keep the weighting they were published with, and the catalogue doesn’t say how much of Neon’s 10-point rise came from changes to the server and how much from changes to the scoring.
What the score doesn’t measure
Manufact’s methodology post keeps curation and score separate, and it puts the limit of the number this way:
doesn’t mean a server is safe or reliable for your task
Manufact, “How Do You Measure MCP Server Quality?”, 23 October 2026
The leaderboard’s FAQ adds that the score doesn’t certify runtime reliability or marketplace approval.

Manufact also has a stake in the ranking. It sells MCP hosting, evals and analytics, and its seed round of US$6.3 million, announced on 12 February, was led by Peak XV according to its company page, which also lists Y Combinator backing at S25. Its own server, mcp.manufact.com, scores 96, third on the board. The methodology post ends with an offer to help teams raise their scores.
Neon’s own documentation carries a warning the score doesn’t. It says the hosted server “grants broad database management capabilities”, recommends MCP for development and testing only, and says never to connect MCP agents to production databases. It asks users to review every action the model requests.
What the server does and how to narrow it
The docs group the server’s tools into 13 categories, including projects, branches, snapshots, SQL, Neon Auth, a Data API, logs, functions and object storage. A documentation category searches Neon’s docs without OAuth. The inspect_database tool runs 15 read-only checks, the same ones as the neon inspect db command, for things such as unused indexes, stalled queries and cache hit rate. SQL tools return up to 100 rows per statement by default, and max_rows raises that to 10,000.
The server connects over Streamable HTTP at mcp.neon.tech/mcp. The older /sse and /message endpoints are retired and return 410 Gone. It supports OAuth and API key authentication.
The access switches sit on the URL. Adding readonly=true leaves SELECT queries and schema inspection and disables writes such as creating branches and running migrations. The roleName switch makes every SQL tool log in as a named Postgres role, so that role’s grants decide what the agent can read. The secrets=none switch keeps credentials out of tool results, and the category and projectId switches narrow the tools and the project. Neon warns that roles made in the console or API, or made without an access level through the CLI or MCP server, belong to neon_superuser and can read all data, so pinning one of those roles limits nothing on reads.

The config and the narrowed URL, from Neon’s docs
A plain OAuth setup, as Neon’s docs give it:
{
"mcpServers": {
"Neon": {
"type": "http",
"url": "https://mcp.neon.tech/mcp"
}
}
}
The same URL with the three switches from the diagram. The role has to exist on every branch you query:
https://mcp.neon.tech/mcp?readonly=true&roleName=app_reader&secrets=none
Manufact reruns a review when a server’s owner asks, and says a score can move when a server’s tools or metadata change, so the 93 is the state of Neon’s server on 20 November. Shamgunov’s post carries no comment beyond the two words.
- Manufact’s board row for Neon shows 17 November as the review date. Neon’s report page shows 20 November at 06:16 UTC, and this article uses the report page. ↩︎
- Manufact’s report header lists 35 tools for Neon, while its documentation review counts 141 discovered tools. The report doesn’t explain the difference. ↩︎
Sources
- Neon MCP Server report, Manufact, reviewed 20 November 2026
- MCP Quality Score Leaderboard, data and analytics, Manufact
- How Do You Measure MCP Server Quality?, Pietro Zullo, Manufact, 23 October 2026
- MCP leaderboard statistics and FAQ, Manufact
- Neon MCP Server overview, Neon docs, last updated 17 November 2026
- Databricks and Neon, Databricks blog, 14 May 2025
- Databricks launches Lakebase, Databricks press release, 11 June 2025
- Manufact raises US$6.3M seed round, Manufact
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.




