The server at /mcp speaks Streamable HTTP and takes the same key as the API. Each tool calls one route and returns its envelope unchanged.
The server is at <base>/mcp (Streamable HTTP). Every tool call needs an API key in the Authorization header; tools/list does not.
claude mcp add --transport http peritus <base>/mcp --header "Authorization: Bearer $PERITUS_API_KEY"{
"mcpServers": {
"peritus": {
"type": "http",
"url": "<base>/mcp",
"headers": { "Authorization": "Bearer <your API key>" }
}
}
}Keep the key out of shared files: most clients read it from an environment variable.
From GET /install.mdAs returned by GET /v1/tools, the same as tools/list. No key needed.
Answer a plain-language question about sanctions status, travel advisory levels, tariffs, or recent changes from stored, cited facts only, with the plan it ran. Use when you have a question in words. Do not use when you already hold entity IDs and fields; call fetch_facts. An abstention is a normal result.
Fetch typed facts with evidence for entities and fields, now or as of a valid and known time. Use when you know entity IDs or exact identifiers and the fields you need. Do not use when you only have a name; call resolve_entity first. Each gap is a typed abstention, never an empty answer.
Find entities whose current facts match a structured filter on a field's value, or free text in source passages, a page at a time, with their facts. Use when you need every entity with some property. Do not use when you have one name; call resolve_entity. No match is a negative result naming the lists searched.
Get the context bundle for one entity: for a country its active advisories, for a listed party its sanctions exposure and identity, each part read from one snapshot. Use when you want everything about an entity you hold. Do not use when you need named fields; call fetch_facts. A missing part makes the bundle partial.
Ask for a field the catalog lacks, in words. A request the catalog already answers returns the field to use and builds nothing; any other is queued for people to build, with a status. Use when no field fits your need. Do not use when you want data; call fetch_facts or search.
Report that an item of one of your responses is wrong: the response ID, the item's fact version, and what is wrong. People review it and the outcome is yours by report ID. Use when a fact you received is wrong. Do not use when a fact is only missing; call request_field.
Resolve names, identifiers, or the known names in a headline to entities with ranked, explained candidates, or a negative result naming the lists searched. Use when you need an entity ID for a name or identifier. Do not use when you already hold the entity ID; call fetch_facts. Scores are not probabilities.
Compare one of your earlier fetch or ask responses with now: each fact added, removed, or changed, with its cause. Use when you hold a response_id and need what changed since. Do not use when you want the change feed in general; call get_changes.
Read typed change events in commit order after a cursor, filtered by entity, field, or domain, and poll with the returned cursor. Use when you track changes over time. Do not use when you want current values; call fetch_facts.
Read an on-demand indexing job that a search queued for days a source had not indexed: its status, its estimated end, and when done the fact versions it wrote. Use when a search returned a PENDING abstention with a job. Do not use when you have no job ID; search again instead.