Skip to content

seismo-mcp

seismo-mcp is a Model Context Protocol server over an operational seismic network. A client — a person, a script, or an autonomous reviewer — calls the same tools with the same contracts. Policy (which tools are visible, which catalogue rows are fenced, how many relocations a session may run) lives on the session, not in the tool.

The server measures. It does not decide what an event is. There is no language model inside the process. Every call is ledgered; a result that is not in the ledger did not happen.

What you get

  • Catalogue origins and picks, by integer id, as the session's view allows.
  • Station inventory and whether a station was delivering, silent, or absent in a time window.
  • Weekly noise baselines and numeric waveform descriptors — never samples.
  • Predicted arrivals, detection probabilities, and a negative-evidence check on delivering stations that have no pick.
  • Relocation, reassociation, a station jackknife, and local magnitude, from a menu of enumerated configurations.
  • A frozen external-bulletin snapshot for teleseism screening and nearby published events.
  • Guided re-picking in a narrow window, and a structured review record whose evidence citations must resolve in this session.
  • Rendered figures for human readers. Pictures are not measurements.

How to read this site

Each tool page opens with the science: what is being measured, how to read a result, and what the method does not imply. The contract below that section is generated from the live registry — input and payload fields, an example, the "does not tell you" paragraph, and any errors or warnings the specification names for that tool. Regenerating the contract does not edit the science.

Identifiers, units, the response envelope, parameter tiers, and the hypothesis rule are on Conventions. Versioning and what this site is not are on About.

Start with relocate if you want the quality bar for a science page, or the tool index to browse by group.