generated: '2026-08-27'
method: derived
source: >-
  mcp/sift-stack-docs-mcp-tools.json (live tools/list from https://docs.siftstack.com/mcp, HTTP 200,
  2026-08-27), https://docs.siftstack.com/documentation/cli/sift-mcp (200, the only published inventory of
  the product server's tools), and openapi/sift-stack-openapi.json.
name: Sift MCP <-> REST tool crosswalk
surfaces:
- {kind: openapi, url: https://docs.siftstack.com/openapi.json, file: openapi/sift-stack-openapi.json, operations: 343, gated: false}
- {kind: grpc, source: https://github.com/sift-stack/sift/tree/main/protos, dir: grpc/, rpcs: 330, gated: false}
- {kind: mcp, id: docs-remote, url: https://docs.siftstack.com/mcp, tools: 3, gated: false, note: 'Anonymous; full inputSchemas captured live.'}
- {kind: mcp, id: product-local, install: 'sift-cli mcp', tools: unknown, gated: true, note: 'Local stdio, API-key authenticated. Not introspected — a full tools/list requires credentials and a local install, so inputSchemas for this server are NOT captured.'}
- {kind: graphql, present: false}
confidence_note: >-
  Two servers, two very different evidence levels. Rows for the docs-remote server are HIGH confidence —
  the tool names and inputSchemas were read from a live tools/list. Rows for the product-local server are
  MEDIUM at best: the tool names are quoted verbatim from Sift's own documentation, but the mapping to REST
  operationIds is by name and semantics, not by introspection, and the docs publish only a partial
  inventory (the gated writes, the additive writes, and two read tools named in passing). No tool was
  invented, and no inputSchema is asserted for the local server.
crosswalk:
- {tool: search_sift, server: docs-remote, category: documentation, rest: [], binding: none, confidence: high, note: 'Full-text search over published docs. No REST equivalent — Sift''s API has no documentation-search operation. DocsService exists in the spec but serves in-product docs resources, not this index.'}
- {tool: query_docs_filesystem_sift, server: docs-remote, category: documentation, rest: [], binding: none, confidence: high, note: 'Read-only shell over a virtualized docs filesystem that also contains /openapi/openapi.json. It is the mechanism by which an agent reads the REST contract, not a projection of it.'}
- {tool: submit_feedback, server: docs-remote, category: feedback, rest: [], binding: none, confidence: high, note: 'Files a docs defect. The only non-read-only tool on the anonymous server, and it has no public REST counterpart.'}
- {tool: create_annotation, server: product-local, category: review, rest: [AnnotationService_CreateAnnotation], binding: direct, confidence: medium}
- {tool: update_annotation, server: product-local, category: review, rest: [AnnotationService_UpdateAnnotation], binding: direct, confidence: medium, gated: true, note: 'Gated behind --allow-destructive. Sift warns it REPLACES tags, metadata and other collections rather than merging them.'}
- {tool: create_report, server: product-local, category: validation, rest: [ReportService_CreateReport], binding: direct, confidence: medium}
- {tool: update_report, server: product-local, category: validation, rest: [ReportService_UpdateReport], binding: direct, confidence: medium, gated: true}
- {tool: create_rule, server: product-local, category: detection, rest: [RuleService_CreateRule], binding: direct, confidence: medium}
- {tool: update_rule, server: product-local, category: detection, rest: [RuleService_UpdateRule], binding: direct, confidence: medium, gated: true}
- {tool: archive_rule, server: product-local, category: detection, rest: [RuleService_ArchiveRule, RuleService_BatchArchiveRules], binding: direct, confidence: medium, gated: true}
- {tool: unarchive_rule, server: product-local, category: detection, rest: [RuleService_UnarchiveRule, RuleService_BatchUnarchiveRules], binding: direct, confidence: medium, gated: true}
- {tool: update_asset, server: product-local, category: catalog, rest: [AssetService_UpdateAsset], binding: direct, confidence: medium, gated: true}
- {tool: update_run, server: product-local, category: catalog, rest: [RunService_UpdateRun], binding: direct, confidence: medium, gated: true}
- {tool: create_test_report, server: product-local, category: validation, rest: [TestReportService_CreateTestReport], binding: probable, confidence: low, note: 'TestReportService exists in the spec; the exact backing operation was not confirmed.'}
- {tool: append_test_measurements, server: product-local, category: validation, rest: [], binding: unknown, confidence: low, note: 'No REST operation with a matching name was found in the spec. Likely a TestReportService or gRPC-side operation; recorded rather than guessed.'}
- {tool: upload_dataset, server: product-local, category: ingest, rest: [DataImportService_CreateDataImportFromUploadV2], binding: probable, confidence: low, note: 'Sift describes this as streaming a computed series back into Sift as a dataset; the backing operation may be a gRPC ingest RPC rather than the REST import.'}
- {tool: data/get_data, server: product-local, category: data, rest: [DataService_GetDataV2, DataService_GetData], binding: direct, confidence: medium, note: 'Named in the "Why the server runs locally" section — downloads channel data locally, typically to Parquet.'}
mcp_only:
- {tool: 'data/sql', reason: 'Runs SQL LOCALLY over data already downloaded with data/get_data. There is no Sift API operation behind it at all — it is compute the MCP server performs on the client machine, and it is the stated reason the product server is local rather than hosted. This is the single most important MCP-only capability in this provider.'}
- {tool: search_sift, reason: 'Documentation search index; no API operation exists.'}
- {tool: query_docs_filesystem_sift, reason: 'Virtualized docs filesystem; no API operation exists.'}
- {tool: submit_feedback, reason: 'Docs feedback channel; no public API operation exists.'}
rest_only_summary:
  operations_total: 343
  operations_with_a_named_tool: 15
  note: >-
    At least 328 REST operations have no tool named in the published MCP inventory. That number is a ceiling
    on the gap, not a measurement of it — the docs publish only the gated and additive-write tools plus two
    read tools, so the local server almost certainly exposes read tools for Assets, Runs, Channels, Reports,
    Campaigns and Families that are simply not enumerated anywhere public. An authenticated tools/list
    against `sift-cli mcp` is the only way to close this, and it needs a customer credential.
  entirely_unrepresented_services:
  - ExportService
  - IngestionConfigService
  - IngestService
  - CalculatedChannelService
  - FamilyService
  - CampaignService
  - WebhookService
  - AutomationService
  - PolicyService / RoleService / UserGroupService (governance)
  - ExternalSyncService
  - DlqErrorsService
coverage:
  mcp_tools_documented: 15
  mapped_to_rest: 11
  mcp_only: 4
  rest_operations: 343
  rest_operations_unmapped: 328
  confidence: {high: 3, medium: 9, low: 3}