Claude Code
claude mcp add serviceproof \
--transport http \
https://api.riproka.com/mcp \
--header "Authorization: Bearer rka1.<your-key>"
Model Context Protocol
ServiceProof is an MCP server, natively. Not a REST API with a translation layer in front of it - the tools your model sees are generated from the same declarations the REST routes are generated from, so they can never describe the platform differently.
https://api.riproka.com/mcpStreamable HTTP · POSTAuthorization: Bearer rka1.<prefix>.<secret>GET /.well-known/oauth-protected-resource844, filtered by what your key may do
Two keys pointed at this endpoint see two different tool lists. The server intersects the
key's scope with the roles of the person who owns it, per call, and offers only what
survives. A model holding a read-only key is not told about delete_customer
and then refused - it never sees it, which is the difference between a
model that behaves and a model that keeps trying.
Destructive capabilities are additionally withheld from any unattended caller. An agent working a queue at 3am has no way to know that cancelling four hundred work orders is not what anyone wanted, so it is not offered the chance.
claude mcp add serviceproof \
--transport http \
https://api.riproka.com/mcp \
--header "Authorization: Bearer rka1.<your-key>"
Add to the MCP config file:
{
"mcpServers": {
"riproka": {
"url": "https://api.riproka.com/mcp",
"headers": {
"Authorization": "Bearer rka1.<your-key>"
}
}
}
}
It is ordinary Streamable HTTP MCP. Any client or SDK that speaks the protocol works - the Python and TypeScript SDKs, an agent loop you wrote yourself, an orchestration platform. There is nothing ServiceProof-specific in the transport.
Full connector guideYou do not write a system prompt describing ServiceProof. These four capabilities exist so the model can find all of that out for itself - and so it stays correct after we ship something new, which a prompt you wrote in March will not.
whoami
Resolves the key to a company, a person and a role mask. The first call any well-behaved agent should make - it is the difference between planning work it can do and planning work it will be refused.
search_capabilities
Searches all 844 capabilities by intent rather than by name, returning each one's
description, required role and blast radius. This is how an agent finds
save_dispatch without you having listed it.
list_documents
Enumerates the published integrator library - the same documents this site serves under /docs.
read_document
Returns a document as markdown. A model that has read the dispatch lifecycle stops guessing which timestamp means travel and which means work.
The corpus those two tools can reach is the published library and nothing else, ever. ServiceProof's internal architecture and reconnaissance notes are not in it - they carry customer names, server addresses and commercial detail, and an integrator's model has no business holding any of that. When an internal document turns out to be genuinely useful to integrators, the fix is to promote that one document into the published set, never to widen the shelf.
MCP, REST, the copilot and the screen all queue behind one executor. There is no path that skips the role check, the audit row or a locking procedure.
A capability that declares a confirmation prompt declares it once. The agent gets the same sentence the screen shows a human - it is not free to be cleverer than a warning somebody wrote on purpose.
A key belongs to a named person's membership of one company. An audit row says who, not "the integration".