Background · Chapter 2October 13, 20265 min read

What already exists: MCP servers and assistants over Allen data

Three kinds of effort have put AI agents in front of Allen data. This chapter covers what each did, how much it was used, and what is still missing.

In short · case study

Most MCP servers wrap an API as tools, and the institute's own server is a broad query wrapper that people use. The unclaimed ground is version-pinned identity and provenance, which the institute's own assistant documents as out of scope.

Tools in the community wrapper
26
over the older Atlas API
Institute MCP releases
57
two servers, 16 months
Tools on the institute's hosted server
25
live, 2026-10-13
Listed MCP projects that are placeholders, forks or abandoned
>50%
ecosystem study, 8,401 projects

Before designing anything, it is worth knowing what has been tried. Three kinds of effort have put AI agents in front of Allen Institute data: a community MCP server, the institute's own MCP servers, and an AI assistant built into the institute's cell atlas. This chapter is about what each did, how much it was used, and what all three leave out.

The figures below were fetched on 2026-10-13 from GitHub, PyPI, npm, the Allen community forum and the live servers. Two claims that did not survive checking are noted below.

Fig. 1Who built what, and when. Bars show how long each effort stayed active.

A community wrapper

In February 2026 an independent developer published allenbrain-mcp, a TypeScript server with 26 tools over the Allen Brain Atlas's older query API. It covered connectivity, ontology, images and downloads. The author announced it on the Allen community forum, called it "still early-stage work", and asked whether there were "existing efforts I should build on instead" (forum post). That was a reasonable and honest way to start.

Seven and a half months later it had drawn little public use (repository). That does not show why. Breadth is one candidate cause, and others are just as plausible: no package to install, no licence, one developer and an older API. The record is also typical. A measurement study of 8,401 MCP projects found "more than 50%" to be placeholders, forks or abandoned, and 21.9% inactive for over a year (Guo et al. 2025). A quiet wrapper is the norm.

The institute's own servers

The Allen Institute for Neural Dynamics has worked on AI access to its experiment metadata for over two years:

  • 2024–2025: a chat assistant (GAMER). It used retrieval-augmented generation on Claude.
  • May 2025: an MCP server. It was carved out of the assistant and reached 26 releases.
  • March 2026: a second MCP server. It followed a major version of the metadata schema, and its first issue asks only to "Reach parity" with the first. It shipped 31 releases by mid-September (PyPI).

That is 57 releases in 16 months. The second server now runs as a hosted endpoint, and on 2026-10-13 it was publicly reachable and listed 25 tools. Staff have filed issues from their own use of it.

This is the strongest evidence against an easy story. In spirit, the institute's server is a broad query wrapper over a database: filters, aggregations, projections and summaries. It is used, and it keeps improving. A study of official MCP servers found that 92% implement their tools as "bare API wrappers" (Mastouri et al. 2025). Wrapping is the normal way to build one.

What the institute's work does show is a direction:

  • Compact results. The server returns "exact leaf-path projections" with an optional "reproducible result marker". Its built-in instructions tell the agent not to fetch records when one aggregation already answers the question.
  • Tools, not resources. In August 2025 the first server replaced resources with tools. In the same change it removed a README line warning that users had to "explicitly ask the agent to use the available resources".
  • A benchmark. The second server has a 15-task benchmark run with two agent models. Its results are not in the repository.
  • A tolerant grading rubric. The earlier chat assistant scored itself 0.75 on 68 questions with an LLM judge whose prompt says "If the numbers are different, that should be okay" (evaluation code). A score is only as strict as its grader.
Two claims that did not survive checking

Early notes for this project described the second server as a redesign prompted by friction with the first. The record shows a schema migration. The only stated reason is "V1 metadata" versus "V2 metadata", and the friction warning had been removed more than six months before the second server began. The same notes counted about 20 releases; PyPI lists 31.

The institute's assistant

In November 2025 the institute added an AI assistant to the Allen Brain Cell Atlas, running Claude Haiku 4.5 on AWS Bedrock. It lets visitors explore and visualise the atlas in natural language. Its technical documentation lists what it does not do: "perform calculations on available data such as cell counts or average gene expression", and "provide detailed provenance or methodology explanations for the data displayed" (documentation).

That is a sound scoping decision for a navigation assistant. In the institute's own words, it also names the gap this project is aimed at.

What this means for the design

  • Do not compete on coverage. A broad wrapper exists, a second broad wrapper adds nothing, and the institute already runs the server that matters for its experiment metadata. The design has seven task-shaped tools (R-10).
  • Follow the institute's direction. That means compact, path-based results that carry a reproducibility marker. The proposal's provenance record (R-1) is that marker, extended to taxonomy, atlas, gene and mapping versions.
  • Grade deterministically. Ids and numbers are checked exactly, and an LLM judge is used only under a written rubric (R-12).
  • Design to be superseded. The institute ships fast. The design uses the institute's own ids (CCF structure ids, taxonomy ids), so its results line up with an official tool.
Fig. 2The cases in this chapter, with what each showed.