Siftable

Update a calendar event

Describe one Vault capability using its non-bearer metadata ID.

Execute one statically reviewed, typed server-brokered operation. Arbitrary URLs, headers, bodies, proxying, and shell commands are unsupported.

List safe metadata for Vault capability handles. No credential or handle bearer material is returned.

Human-only admission decision. MCP agents cannot confirm durable code memories; review the candidate in the Siftable human interface.

Hard-delete a code memory. Prefer reject/supersede for curation; delete only when explicitly required.

Scan for likely-stale code memories (missing evidence, retired repo) and mark them needs_review. Never deletes.

Human-only curation decision. MCP agents cannot edit admitted code memories; store a new candidate or mark the memory needs_review.

List append-only curation/lineage audit events for a code memory

Get one code memory including provenance, evidence, confidence, authoring agent, and lifecycle state

List stored facts with provenance and lifecycle. Rejected/superseded excluded by default.

Mark a likely-stale or contradicted code memory for human review. Never deletes.

Human-only rejection decision. MCP agents cannot reject durable code memories; review the candidate in the Siftable human interface.

Search curated code-memory facts previously stored in Siftable; this does not inspect or search repository files. Default retrieval excludes rejected/superseded and ranks confirmed above unreviewed. Pass scopePaths for path proximity boost.

Store a durable codebase fact. Agent-authored stores should include authoringAgent, captureReason, repositoryId, and path/evidence refs. Secrets and raw source dumps are refused.

Human-only supersession decision. MCP agents cannot hide an admitted fact; store a replacement candidate or mark the memory needs_review.

Explain the current MCP workspace, project, repository, transport, version, and effective authorization context. Hosted MCP never invents a client checkout; pass an explicit repository/project ID or clientRootHint when needed.

Apply only approved operations from an existing durable CRM import plan. The server preserves per-operation receipts and safely replays completed work; this tool cannot bypass approval.

Record a human approval decision for all or selected stable operation IDs in an existing CRM import plan. This does not apply canonical CRM writes.

Create or reuse a durable review-only CRM import plan from an exact @siftable/tabular-import projection. Supports CSV, TSV, XLS, and XLSX projections serialized within 8 MiB, with 32 sheets, 1,000 columns per sheet, and 25,000 rows. It preserves provenance and duplicate/conflict decisions and does not write canonical CRM records. HTML must be normalized to CSV or XLSX first.

Resume and inspect one durable CRM import plan by stable plan ID, preserving server operation IDs, provenance, decisions, conflicts, and apply receipts.

Aggregate dataset records with grouped metrics: count, avg, sum, min, max, distinct_count, median, stddev, percentile, ratio. percentile uses a 0-1 fraction and continuous linear interpolation. stddev is population standard deviation. Field-based aggregates exclude null values. Supports HAVING clause.

Generate grounded natural-language insights for a dataset slice using aggregates, facets, and evidence rows.

Bucket a numeric or date field into ranges and compute metrics per bucket. Provide boundaries or bucketCount for auto-bucketing.

Compare metrics across different segments of a categorical field side-by-side with deltas.

Compute derived numeric columns using arithmetic plus safe functions like lag, diff, pct_change, rolling_avg, rolling_sum, and ratio.

Create a dataset

Export dataset records as CSV. Supports filtering and sorting.

Get value distributions (facets) for dataset fields — useful for understanding data shape and filtering options

Description is built at runtime.

Join a dataset to itself with alias-scoped fields such as left.Close and right.Close.

List datasets

Materialize a derived result into a new scratch dataset.

Mutate dataset records (create/update/delete)

Validate and normalize a lightweight chart payload from a derived dataset result.

Analyze dataset quality with structured missing-value metrics, actionable issues, and bounded repeated-value observations. Repetition is not treated as a defect unless the field is declared unique.

Query dataset records by filters/sorts. Natural-order pagination uses seek cursors; custom-sort pagination uses offset cursors and is not snapshot-stable across concurrent changes.

Rank dataset records by explicit sorts or a weighted numeric formula.

Modify dataset schema (add/update/delete field)

Summarize dataset identity, schema, lifecycle, formula/relation metadata, and up to five sample rows

Analyze a dataset slice as time series data with lag, rolling windows, drawdown, volatility, correlation, and optional pivoting.

Traverse the entity relationship graph centered on any entity (task, note, person, project, dataset, etc.). Returns connected nodes and edges. A node's status slot carries the record's own status where it has one (tasks: inbox/next_action/in_progress/waiting_for/completed/archived; notes/datasets: active/archived; projects: their status); other types show a type-specific label (organizations: industry, people: relationship, news: published date).

List reviewed local execution adapters, including honest contained, bounded_only, and unsupported threat disclosures.

Request a human-approved ephemeral grant for a separately prepared runner public key. The human must verify the displayed key fingerprint and executable digest; this is approval-based trust, not hardware attestation. MCP cannot approve, redeem, or receive plaintext credentials.

Inspect safe status and containment metadata for an execution grant. No bearer or credential material is returned.

Search the active Siftable capability catalog by natural-language intent, including tools that are unavailable to the current client. Returns separate authorization and runtime availability, schema digests, and exact recovery guidance. Retired source capabilities return recovery cards.

Git blame (local)

List the review-only CRM candidates, source identities, and duplicate-match state produced by one lead discovery run.

Shortlist, reject, or restore a lead discovery candidate. This records review state only; it does not import a canonical CRM record or enroll a prospect.

Description is built at runtime.

Prepare an exact, bounded company-discovery quote without dispatching provider work. Returns an opaque prepared-query ID, digest, expiry, warnings, and maximum charge; no execution token is exposed.

Run the exact server-held prepared company-discovery contract after explicitly approving its quoted maximum charge. The opaque prepared-query ID is single-use and no execution token is accepted or returned.

Run one bounded, paid Crustdata company search and save the returned organizations as review-only CRM candidates with source provenance and duplicate matching. This never creates canonical CRM records or prospects. The response reports the maximum and actual charge; billing and one-use provider authority are enforced by the Siftable backend exactly as in the app.

Create or recover the import plan bound to one exact reviewed shortlist digest. This does not approve or apply the plan.

Recover the immutable import-plan receipt bound to a discovery run and exact shortlist digest.

Get persisted discovery-origin evidence for an imported organization in the current personal or workspace scope.

Recover durable status, typed failure or hold details, and the stable receipt for a company-discovery run after a disconnect or duplicate submission.

Preview an exact note snapshot, then apply only that snapshot with confirm, previewId, expectedSnapshotDigest, and a stable idempotencyKey

Create a note

Delete a note

Get a single note with full content

List notes

Search notes. Full-text matches (PostgreSQL websearch syntax over title and body) rank before semantic neighbors, then literal-substring fallbacks; within each group higher score first, so scores are comparable within a source, not across sources.

Update a note

Run declarative action on an ontology object

Find ontology objects by object type and property filters

Traverse graph links for an object record

Preview an exact organization snapshot, then apply only that snapshot with confirm, previewId, expectedSnapshotDigest, and a stable idempotencyKey

Create a new organization/company

Delete an organization (unlinks people but does not delete them)

Check whether organization enrichment providers are configured and return the current standard and ultra app prices. This does not dispatch providers or charge the account.

Run the same paid, idempotent organization enrichment preview used by the app. It saves source evidence and returns reconciliation proposals with billing evidence. The current app contract is preview-only and does not permit canonical CRM application.

Search organizations/companies

Update an existing organization/company

Preview an exact contact snapshot, then apply only that snapshot with confirm, previewId, expectedSnapshotDigest, and a stable idempotencyKey

Search contacts

Create a new person/contact

Delete a person/contact

Apply selected profile fields and facts from a completed, settled person enrichment run to the same CRM person record. The backend rehydrates stored proposals by field and fact ID; callers cannot substitute free-form values. This step does not dispatch or charge another provider.

Check whether a CRM person has enough public professional identity data for enrichment and return the current app prices. This does not dispatch providers or charge the account.

Run the same paid, idempotent public-professional person enrichment product used by the app. It returns sourced profile and fact proposals and billing evidence; it does not automatically overwrite the CRM record.

Update an existing person/contact

Archive a project

Create a project

Get project context (tasks, notes, members, signals)

List projects

Get the canonical CSN planning snapshot for a project

Recompute the canonical CSN planning snapshot for a project

Update a project

Inspect the canonical read-only contact plan for one person, including lifecycle state, communication policy, collisions, diagnostics, and exact source-ledger locators. This tool cannot mutate, schedule, approve, enroll, or send.

Persist a durable repository-to-project identity after context_current returns an unlinked repository and the caller selects a project candidate. This stores metadata only; it does not upload, index, search, or delete source code.

Inspect how to acquire an unavailable Siftable capability without interrupting the MCP session. On Streamable HTTP OAuth connections inspect returns a short-lived approval link (elevation) that widens this connection in place; otherwise it returns the terminal OAuth/reconnect action. Challenge remains available only for clients that explicitly handle HTTP 403 insufficient_scope responses.

Discover the data model: entity types, link types, core entity schemas (fields), dynamic datasets, and ontology object types. Datasets are summarised (title, ID, rows, field count) unless includeDatasetFields is true; dataset_summarize returns one dataset's field definitions.

Preview an exact task snapshot, then apply only that snapshot with confirm, previewId, expectedSnapshotDigest, and a stable idempotencyKey

Mark task complete

Create a CSN coupling edge between tasks in the same project

Delete a CSN coupling edge from a task

List CSN coupling edges for a task

Create a human planning task. Executable agent work is created separately with work_item_create (work:write), optionally linked to this task by taskId; if that tool is not callable on this connection, request_capability names the exact next step.

Delete a task

Get a single human planning task with its complete derived state and the executable work linked to it (when this connection can read work items; otherwise work_item_list with taskId needs work:read and request_capability names the next step).

List human planning tasks with stable IDs and complete derived state. status=in_progress is accepted as an alias for phase=in_flight. Agent execution state lives on separately authorized work_item records.

Update CSN planning fields on a task

Update a human planning task — supports brief fields, planning state, and goal linkage. Agent execution should be updated through work_item tools.

Upload a document (PDF, Markdown, or text) into Knowledge as a note

List Vault audit events (requires vault:audit:read). Returns safe audit metadata only.

Store a new encrypted secret in the vault (requires vault:manage)

List vault entries (requires vault:metadata:read). Returns metadata only — never decrypted values.

Request human approval for one exact plaintext-at-rest destination using a separately prepared runner public key. MCP receives safe request metadata only and cannot approve, redeem, or write a path.

Inspect safe metadata and status for a Vault materialization request. No artifact, ciphertext, or plaintext is returned.

Search vault entries by name, slug, or description (requires vault:metadata:read). Returns metadata only.

Update vault entry metadata (requires vault:manage)

Get the project default required gate for work-item dependencies

Set the project default required gate for work-item dependencies

Claim the next queued executable work item with a transactional lease

Monotonically advance a durable mailbox cursor only after every event through lastSequence has been processed. Only lastDeliveredSequence from list/wait is safe when a page is truncated.

Read up to 10 ordered work-item mailbox events after a sequence. Reading never acknowledges delivery.

Append an agent-authored message or artifact publication to a work-item mailbox. Observers cannot send, and this does not change work lifecycle state.

Wait for up to 10 ordered work-item mailbox events for a stable consumer. Omit afterSequence to resume from its durable cursor. Waiting never acknowledges delivery.

Add a contributor, reviewer, or observer to a work-item mailbox. This cannot create or replace the coordinator; coordinator authority remains the work-item lease.

List up to 100 participants in a work-item mailbox as an explicit agent participant. This does not change the coordinator lease.

Remove an assignable participant from a work-item mailbox. This cannot remove or transfer the lease-derived coordinator.

Create executable agent work with an optional canonical Work Contract V1, authoritative UUID dependencies, and parent human planning task linkage.

Atomically replace the authoritative dependency edges for a work item

C2: Revise queued, unclaimed work while preserving its UUID and dependency edges. Do not use after claim or execution; wait for a queued unclaimed state or create follow-up work. Example: raise queue rank or refine acceptance criteria while the item is queued and unclaimed. Counterexample: do not change status, dependencies, or an active worker lease with this capability. Execution-contract edits supersede the active verification plan and reset verification evidence; queue-rank-only edits do not. The same Idempotency-Key and payload replays the original receipt; a different payload with that k…

Get executable agent work item details, authoritative dependencies, and derived claimability

List executable agent work queue items. Active work only by default (done/failed/cancelled are hidden unless status names one or includeTerminal is true). Ordered by status (running, claimed, queued, blocked, needs_review), then queueRank descending (higher = earlier; 0 is the default; negative ranks are deferred and sort after 0), then oldest first. Use taskId to see execution linked to a human planning task.

List immutable work-item edit receipts with actor attribution, before/after snapshots, and field-addressable diffs.

Submit externally executed evidence for an exact active plan version and stable step ID. This never executes a command or revises criteria.

List immutable verification plan versions, superseded evidence, and legacy unbound evidence

Inspect the active immutable verification plan, stable step IDs, latest attempts, and missing coverage

Release a principal-bound hosted work lease and invalidate its opaque session handle.

Claim executable work through a principal-bound hosted session. The claim token remains server-side; use the returned sessionHandle for later lifecycle calls.

Advance work through a principal-bound hosted session without accepting or returning a raw claim token.