io.github.cyanheads/cern-opendata-mcp-server

@cyanheads/cern-opendata-mcp-server

Search CERN Open Data, fetch records, files, analysis environments, CMS good-run lists, HLT paths via MCP. STDIO or Streamable HTTP.

7 Tools • 1 Resource

Public Hosted Server: https://cern-opendata.caseyjhand.com/mcp

Overview

Particle-physics data from the CERN Open Data Portal: collision and simulated datasets, analysis software, and documentation from ALICE, ATLAS, CMS, LHCb, and more. Search with live facet counts, open records with license and citation, list data files, assemble a record's analysis environment, and look up CMS good-run lists and trigger paths. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above.

Tools

Resources

Tool-only clients get the same data from cern_opendata_get_records.

Capability reference

cern_opendata_search_records tool

• Optional query (an OpenSearch query_string, up to 500 characters) plus OR-list filters type, experiment, category (physics category, Higgs Physics::Standard Model), keywords, collision_energy, collision_type, file_type, availability, collection, and the LHCb magnet_polarity, stripping_stream and stripping_version, each an array of values; year_from/year_to and min_events/max_events bound the data-taking year and event count
• sort (bestmatch, mostrecent for newest date_published first, title, title_desc), limit 1–50 (default 10; a larger value, up to 10,000, is served at 50 with a notice) and page from 1, counting pages of the limit applied; page × limit past 10,000 fails as page_window_exceeded
• Compact hits with recids, plus thirteen live facets that each ignore their own filter (type and category with up to 10 secondary values each, and subtypes_other_count or subcategories_other_count when the portal cuts that list); applied_filters echoes what ran, and values outside the verified vocabulary appear under unrecognized_values

cern_opendata_get_records tool

• ids: up to 100 recids, DOIs, CMS dataset paths (/Primary/Era/TIER) or documentation slugs, mixed in one array; the first 20 distinct are looked up and the rest returned in deferred
• Each record carries a license with its basis (record, cern_terms_default, not_stated) and, when it has a DOI, a ready citation
• Documentation and news bodies come in slices of up to 30,000 characters: body_offset with that one id reads on from the body_next_offset the last slice returned
• Each response stays within 64,000 bytes, except that the first record always comes back whole: records past the budget are left out and listed in deferred, ahead of any ids past the first 20, to pass back as ids
• Records also return what they state of a variables dictionary (name, type, unit, description), a physics category, pile-up (pileup_html, with the pile-up datasets under links), keywords, and the LHCb magnet_polarity and stripping stream and version
• Unresolved identifiers land in missing with interpreted_as and guidance instead of failing the call; file lists come from cern_opendata_list_files

cern_opendata_list_files tool

• recid required; without index, returns the record's file indexes and regular files, and with an index key, that index's files, read without the rest of the record
• limit 1–500 (default 50; a larger value, up to 10,000, is served at 500 with a notice), continued with next_cursor; each file carries xrootd_uri, size_in_bytes, checksum and availability, an https_url when one can be built from its key or EOS path, and key or filename as the portal states them, and each index a uri_list_url listing every XRootD URI in it
• Files marked on demand sit on tape and must be requested on the record's portal page first; an umbrella record with no files of its own returns its child recids under children

cern_opendata_get_analysis_env tool

• recid required; software carries the record's own container images, CMSSW release, global tag and environment recid
• environment_records (condition, VM, validation) for the record's run periods and example_software that declares it works with the record, up to 50 between them; guides quotes the linked section of the first two portal guides, each capped at 12,000 characters, with the [label]: url reference definitions the quote links through from elsewhere in the guide, and a cut names the cern_opendata_get_records body_offset that reads on from it
• Always separately_licensed: true; linked records or guides that can't be read leave a notice instead of failing the call

cern_opendata_get_validated_runs tool

• Exactly one of recid (a CMS collision dataset or a validated-run list) or run_period (Run2012B; 2012B also matches); variant full or muons_only; run_min/run_max; limit 1–2000 (default 200; a larger value, up to 10,000, is served at 2000 with a notice)
• A dataset recid bounds the runs to the dataset's first and last listed run, echoed in run_bounds; when several lists match, matched_lists names them and no runs are read
• Each run carries lumi_sections and lumi_ranges; list.https_url downloads the whole list file. CMS only: other records fail as no_validated_runs

cern_opendata_search_trigger_paths tool

• path: an exact name (HLT_IsoMu24, AlCa_EcalPi0) or a prefix with one trailing * (HLT_IsoMu*); a name without the HLT_ prefix is matched against record path names in the case given and also searched with HLT_ added, and a _v<n> version suffix is dropped; optional year, limit 1–50 (default 10; a larger value, up to 10,000, is served at 50 with a notice) and page
• Each per-year record is parsed into the primary datasets its title names, first_seen, last_seen, per-version run ranges with their l1_seed, and HLT menu record links; parsed: false marks a record to read from its abstract_html
• CMS open data from 2011–2016 only

cern_opendata_list_reference tool

• Optional topic: experiments, record_types, collision_energies, collision_types, file_types, availability, categories, lhcb, identifiers, query_syntax, licensing or run_periods; omit it for every table
• Static and offline; categories, lhcb and run_periods are dated snapshots, while the search facets and cern_opendata_get_validated_runs read the live portal

cern-opendata://record/{recid} resource

• One record by recid (6004, or a prefixed recid such as atlas-160006; leading zeros ignored) as application/json, in the cern_opendata_get_records record shape: metadata (variable dictionary, physics category, pile-up and LHCb run conditions included), license and citation, without file lists
• recid comes from cern_opendata_search_records; reads carry a 15-minute public cache hint

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

CERN Open Data-specific:

• Keyless and read-only; it never stages tape files or writes anything
• One shared pacer at 50 requests a minute, under the portal's published 60 per client IP, and one 50-second deadline per call across queue wait and retries
• Filter values canonicalized against the portal's verified vocabulary (13 tev → 13TeV, lhcb → LHCb, Pb-Pb → PbPb, dataset/collision → Dataset::Collision); unknown values are sent as given and flagged
• Tape-resident (ondemand) records are included in every search and lookup, where the portal otherwise drops them silently
• File manifests, and file indexes read on their own, are cached for 15 minutes, so paging through a record's or an index's files costs one read
• Records cern_opendata_get_records and the record resource resolved are cached for 15 minutes (up to 8 MiB), so fetching a budget-deferred record or following body_next_offset costs no further portal request

Agent-friendly output:

• Provenance on every response: portal_url on each hit and record, a license with its basis, a DOI citation, and applied_filters or effectiveQuery echoing what ran
• Graceful partial results: unresolved ids land under missing with guidance, and unreadable linked records or guides in cern_opendata_get_analysis_env surface as a notice rather than a failure
• Discriminated outputs: kind, license.basis, interpreted_as, scope, variant, run_bounds.source and parsed let callers branch on data, not string parsing
• Portal text kept as data: titles, descriptions, guide sections and file names are fenced or escaped in content[] and relayed as received (HTML in _html fields) in structuredContent

Data and licensing

Portal metadata and datasets are CC0 under the CERN Open Data Terms of Use. Software, container images, documentation and guide code are licensed separately, per record (software is commonly GPL). cern_opendata_get_records reports each record's license and its basis: record when the record states one, cern_terms_default for a dataset that states none (CC0), and not_stated otherwise.

CERN asks reusers to cite each dataset's DOI in applications and publications; cern_opendata_get_records returns a ready citation for every record with a DOI.

This server is an independent project and is not affiliated with or endorsed by CERN.

Known limitations

• 60 requests a minute per client IP. The portal publishes this limit; the server paces itself to 50 a minute, and a call that cannot start within its deadline fails as rate_limited with retryAfter. A hosted deployment shares one budget across every user behind its egress IP. cern_opendata_get_analysis_env and cern_opendata_get_validated_runs cost 2–4 requests each.
• 10,000-result window. Search and trigger-path paging reach only the first 10,000 matches; narrow deeper result sets with filters.
• Facet lists are partial. Terms facets return the first 10 values alphabetically (file_type up to 100), with the rest counted in other_count, and each type and category bucket its first 10 secondary values, with the rest counted in subtypes_other_count or subcategories_other_count (an unfiltered search hides Supplementaries::Trigger this way). These are summed counts of values, not record counts; cern_opendata_list_reference lists every value, and a Primary::Secondary filter's totalCount counts its records.
• Tape-resident files. Files with availability on demand must be requested on the record's portal page before download; staging them is a write and out of scope. A record whose availability is ondemand lists none of its files through the API, so cern_opendata_list_files reports only the count and size its metadata states.
• CMS-only run lists and triggers. Trigger records cover 2011–2016, and trigger detail is limited to what each record's abstract states. cern_opendata_search_trigger_paths returns no prescale factors: they are in each HLT menu record's configuration file (its PrescaleService block, listed by cern_opendata_list_files with the menu_recid), and the CMS luminosity records (collection CMS-Luminosity-Information) log which prescale column applied by run and luminosity section. Muons-only run lists do not exist for Commissioning2010, Run2010B or the 2011 ReReco list.
• Glossary entries are not served. The portal's glossary links answer 404, so search excludes them.

Getting started

Public Hosted Instance

A public instance is available at https://cern-opendata.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "cern-opendata-mcp-server": {
      "type": "streamable-http",
      "url": "https://cern-opendata.caseyjhand.com/mcp"
    }
  }
}

Every caller of the hosted instance shares one portal request budget of 50 requests a minute. For sustained use, run your own instance.

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "cern-opendata-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/cern-opendata-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "cern-opendata-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/cern-opendata-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "cern-opendata-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/cern-opendata-mcp-server:latest"]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No API key or account: the CERN Open Data Portal is public.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/cern-opendata-mcp-server.git

1. Navigate into the directory:

cd cern-opendata-mcp-server

1. Install dependencies:

bun install

1. Configure environment:

cp .env.example .env
# optional: adjust transport, logging, or telemetry settings

Configuration

The server has no settings of its own; these framework variables apply.

See .env.example for the common framework overrides.

Running the server

Local development

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging and ctx.enrich for notices and paging context
• Register new tools and resources in the barrels in src/mcp-server/*/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.