Atlas · MCP

The Atlas for AI assistants

An AI assistant that supports MCP (Model Context Protocol) can query the IsoFind Atlas directly: find a deposit, read its values and their sources, compare a sample, convert a value between standards, search fractionation factors, compose an IGDB database, read an IGL model. No account, on the same terms as the website.

Server address: https://atlas.isofind.tech/mcp
Transport: MCP Streamable HTTP, no authentication
Tools: 18, all read-only
Quotas: Those of the website, plus limits of its own for assistants

01Connecting an assistant

Add the server to the assistant as a remote MCP server, or "custom connector", with the address https://atlas.isofind.tech/mcp. It needs no key and no account. Where this setting lives depends on the assistant and changes between versions: look in its settings for connectors or MCP servers.

In Claude Code, from the command line:

claude mcp add --transport http isofind-atlas https://atlas.isofind.tech/mcp

In editors and clients that read a JSON configuration, the most common form is the following; key names vary from one program to another:

{
  "mcpServers": {
    "isofind-atlas": {
      "type": "http",
      "url": "https://atlas.isofind.tech/mcp"
    }
  }
}

Once the server is added, just ask: "Which δ⁶⁵Cu values does the Atlas give for Chilean porphyry copper deposits, with their sources?", "Convert δ⁶⁵Cu = 0.50 ‰ measured against ERM-AE647 to the NIST SRM 976 scale", "Write an IGL model of Cr(VI) reduction using a factor from the Atlas". The assistant chooses which tools to call.

Opened in a browser, the address returns a short description of the server and the list of its tools.

02What the assistant can do

Tool	What it returns
Atlas records
`search_records`	ore deposits and contaminated sites found by name, country, region or element, with their id, without values
`get_record`	a complete record: every value with its system, uncertainty, range, number of analyses, standard, mineral and source publication with its DOI
`count_compatible_sources`	how many records are compatible with the measured values of a sample; the full ranking remains part of Atlas access
`list_isotope_systems`	the isotope systems present and their exact notation
`get_distribution`	the distribution of one system by deposit family, country or mineral: quantiles, counts, curves
`get_data_coverage`	counts per element and per family, country or mineral: where the Atlas has data and where it has gaps
Fractionation
`get_fractionation`	the fractionation data of one element, by process
`search_fractionation_factors`	fractionation factors filtered by element, ratio, mechanism, context, regime or text, with their citation
Standards
`list_standards`	the standards of one system, their value on the common scale and the publication of each offset
`convert_standard`	a δ value moved from one standard to another with the exact composition formula, from published offsets only
IGDB databases and IGL models
`list_igdb_databases`, `get_igdb_database`	the per-system IGDB databases, their SHA-256 digest and their text
`compose_igdb_database`	an IGDB database composed from filters, with its permanent address
`search_igl_models`, `get_igl_model`	the published models of the library, their source text, databases and citation
`get_igl_reference`	the reference of the IGL language, by topic, for writing or explaining a model
Versions
`list_atlas_versions`	the frozen, citable versions of the Atlas
`get_recent_changes`	the feed of changes: publications added, new records, corrections, models

No tool writes anything: the server offers no contribution, publication or modification.

03Quotas, use and citation

The server is there to answer questions, not to copy the Atlas. On top of the website quotas, counted per network address over 24 hours (article 9 of the Atlas terms), it adds its own limits on reading records and fractionation data: 200 distinct objects per network per 24 hours, 60 reads per network per hour, 30 distinct records per conversation. An online assistant calls from its provider's addresses: the per-network limit caps what that assistant reads for all of its users together. Searches return identifiers, not values. As on the website, the full ranking of compatible sources needs Atlas access, whose session token can be sent in the X-Atlas-Token header.

IsoFind reserves text and data mining rights on the Atlas values (EU Directive 2019/790, article 4; Atlas terms, articles 7 to 9), subject to the exceptions the law makes mandatory. Every response that carries values says so in plain text, and the TDM-Reservation: 1 header (TDMRep protocol) says so to machines: the values are provided to answer the user's question, quoted with their source; they may not be used to train, fine-tune or evaluate AI models, to build datasets, or to copy the Atlas, in whole or in part. A network or an assistant that reads in bulk can be suspended.

These measures limit what an assistant can extract and make the reservation explicit; they do not prevent an assistant's provider from keeping what the assistant received during a conversation.

Every value returned carries its source. The server asks assistants to cite the publication of each value, with its DOI where there is one, to name the IsoFind Atlas, and to give the Atlas version when reproducibility matters. It also reminds them that a contaminated site is not a deposit signature, that a range is not a mean, and that conversions between standards use published offsets only.

The IGDB databases, under the CC-BY 4.0 licence, and the library models, published under their authors' licences, are not covered by the reservation.

04Technical details

Streamable HTTP transport: one address, JSON-RPC 2.0 messages over POST, one JSON response per request, no SSE stream.
Protocol versions: 2026-07-28, stateless (with server/discover), and 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05, with the initialize handshake.
Tools are annotated as read-only and idempotent. A result rarely exceeds a few tens of kilobytes; beyond 120 kB it is truncated, with a message asking to narrow the request.
A tool error (unknown record, quota reached) comes back as a result the assistant can read, not as a protocol failure.

The server counts its calls per assistant (from the name the software gives itself), per tool and per day. Nothing identifies a person: the network address is only kept as a fingerprint, to count distinct networks.