pipefy-introspection

Introspection & Raw GraphQL

Schema discovery, organization info, and a fallback executor. 7 MCP tools.

This is Tier 2 in the resolution strategy: when a dedicated MCP tool fails or doesn't exist, use introspection to understand the API, then execute_graphql to run the operation directly.

Tier 1: dedicated MCP tool exists — use it. Tier 2: use introspection + execute_graphql (this skill). Tier 3: direct curl/httpx fallback — see skills/api-troubleshoot/pipefy-api-fallback/SKILL.md.

Tools

The include_parsed flag

execute_graphql, introspect_type, introspect_mutation, introspect_query, and get_organization all accept an optional include_parsed: bool (default false).

• Default (`false`): response is { success, result } where result is the raw GraphQL JSON as a string.
• `true`: response includes both result (the raw JSON string) AND data (the parsed dict). Drill into data programmatically; keep result to forward verbatim.

Use include_parsed=true whenever you plan to read nested fields (e.g. iterating over phases[].fields[]). Leave it off for one-shot reads where the raw string is sufficient.

max_depth (introspect_type / query / mutation)

MCP tools accept max_depth (default 1). CLI: --max-depth.

• `1` — type/field info only (no inlined sub-types).
• `2+` — resolves referenced input/output types inline (resolvedType), so one call can replace introspecting the mutation then each input type separately.

Example (CLI):

pipefy introspect mutation createCard --max-depth 2 --json

Example (MCP):

introspect_mutation mutation_name="createCard" max_depth=2 include_parsed=true

Scalars (ID, String, Int, …) are never expanded.

kind on search_schema

Optional filter: OBJECT, INPUT_OBJECT, ENUM, SCALAR, INTERFACE, UNION.

search_schema keyword="automation" kind="INPUT_OBJECT"

pipefy introspect schema search automation --kind INPUT_OBJECT --json

When to use introspection

• A dedicated tool returned an error and you need to understand why — introspect the input type to check argument names/types.
• Before creating fields: introspect_type('CreatePhaseFieldInput') to discover valid type enum values.
• Before using extra_input: introspect the corresponding input type to find optional keys.
• Unknown mutation signature: introspect_mutation('createSomething') before execute_graphql.
• Schema exploration: search_schema('automation') to find related types and inputs.

When to use execute_graphql

• No dedicated MCP tool exists for the operation.
• A dedicated tool failed and you've used introspection to understand the correct payload.
• Ad-hoc queries like resolving an org UUID via pipe(id: $id) { organization { uuid } }.
• Complex nested queries that no single tool covers.

Always prefer dedicated MCP tools. They validate inputs, handle pagination, and format errors consistently. execute_graphql is the fallback when dedicated tools can't solve the problem. Queries are ungated; MCP mutations need the two-step confirmation_token.

Steps — discover a mutation signature

1. Search for the mutation by keyword:

   search_schema keyword="label"

1. Get the full mutation signature:

   introspect_mutation mutation_name="createLabel"

1. Discover input type fields:

   introspect_type type_name="CreateLabelInput"

1. Execute the mutation (CLI):

   pipefy graphql exec --query "mutation …" --vars '{"input":{…}}' --yes --json

Mutations: the CLI exits with code 2 unless --yes is passed (guardrail for agents and scripts). It also exits 2, with or without --yes, when the document is too deeply nested to parse: nothing is sent, because a document that cannot be classified could carry an unconfirmed mutation. execute_graphql refuses the same document with an error payload.

1. Execute the mutation (MCP): two-step. The first call returns a preview with confirmation_token and does not mutate. The preview names the mutation; it does not claim the write is irreversible, because this server cannot tell create from delete. Resend the call unchanged with confirm=true and the token; if the document changed, the response is a fresh preview whose token is bound to the new document.

   execute_graphql query="mutation CreateLabel($input: CreateLabelInput!) { createLabel(input: $input) { label { id name } } }" variables='{"input": {"pipe_id": 67890, "name": "Urgent", "color": "#FF0000"}}'

Then after the preview:

   execute_graphql query="mutation CreateLabel($input: CreateLabelInput!) { createLabel(input: $input) { label { id name } } }" variables='{"input": {"pipe_id": 67890, "name": "Urgent", "color": "#FF0000"}}' confirm=true confirmation_token="<token from preview>"

Common fallback recipes

Ready-to-use patterns for situations where dedicated tools are insufficient.

Recipe 1 — Discover valid field types for create_phase_field

introspect_type('CreatePhaseFieldInput')

Look for the type field; it references an enum. Introspect the enum to get all valid values.

Recipe 2 — Get full behavior config of an AI agent

get_ai_agent returns behavior headers only. To inspect the full config (event_params, actionParams, actionsAttributes):

execute_graphql query='query($uuid: ID!) { aiAgent(uuid: $uuid) { uuid name instruction behaviors { id name active event_id event_params { to_phase_id triggerFieldIds fromPhaseId } action_params { aiBehaviorParams { instruction referencedFieldIds actionsAttributes { name actionType referenceId metadata { destinationPhaseId pipeId fieldsAttributes { fieldId inputMode value } } } } } } } }' variables='{"uuid":"<agent-uuid>"}'

Recipe 3 — Find a card by title (not possible with find_cards)

find_cards only searches custom field values. To search by title:

execute_graphql query='query($pipeId: ID!, $first: Int) { cards(pipe_id: $pipeId, first: $first) { edges { node { id title current_phase { name } } } } }' variables='{"pipeId":"<pipe-id>","first":50}'

Filter by title client-side. For large pipes, paginate with after.

Recipe 4 — Discover what extra_input accepts for any mutation

When a tool accepts extra_input (e.g. create_automation, update_label), discover all optional keys:

introspect_mutation('createAutomation')     # find the input type name
introspect_type('CreateAutomationInput')    # see all inputFields

Compare with the tool's primary arguments to know which keys are additive via extra_input.

Recipe 5 — Discover organization IDs

To answer "which organizations do I have access to?" with nothing in hand, call list_organizations — it needs no id and returns each org's id, uuid, name, and your role. That is the entry point; reach for the GraphQL fallbacks below only when you already have a pipe.

When the user only has a pipe ID and needs its organization_id:

execute_graphql query='query($id: ID!) { pipe(id: $id) { organization { id uuid name } } }' variables='{"id":"<pipe-id>"}'

Recipe 6 — Update a select field's options after creation

create_phase_field does not accept options. Create first, then update. MCP mutations are two-step: preview, then confirm=true plus confirmation_token. Resend the call unchanged with confirm=true and the token; if the document changed, the response is a fresh preview whose token is bound to the new document. CLI --yes can stay one-shot.

execute_graphql query='mutation($id: ID!, $options: [String!]) { updatePhaseField(input: { id: $id, options: $options }) { phase_field { id label options } } }' variables='{"id":"<field-id>","options":["High","Medium","Low"]}'

Then after the preview:

execute_graphql query='mutation($id: ID!, $options: [String!]) { updatePhaseField(input: { id: $id, options: $options }) { phase_field { id label options } } }' variables='{"id":"<field-id>","options":["High","Medium","Low"]}' confirm=true confirmation_token="<token from preview>"

Recipe 7 — Check phase transition rules

When move_card_to_phase fails with "not a valid target phase":

execute_graphql query='query($id: ID!) { phase(id: $id) { id name cards_can_be_moved_to_phases { id name } } }' variables='{"id":"<current-phase-id>"}'

Returns the valid destination phases from the current phase.

Optional schema cache

For long-running agent sessions, the MCP can reuse the fetched GraphQL schema across requests instead of re-introspecting on every call. Enable via the gql_reuse_fetched_graphql_schema setting (env or settings file). Off by default. After a breaking Pipefy schema change, the process must be restarted to pick up the new schema. Single-session agents rarely benefit — leave it off unless you measure real improvement.

Success criteria

• introspect_type returns the complete field list for the input type.
• execute_graphql returns the expected data without errors.

Failure modes

• `introspect_type` returns `null` — type name is case-sensitive; try PascalCase (e.g., CreateLabelInput, not create_label_input).
• `search_schema` returns many hits — case-insensitive substring matching; broad keywords like "card" flood results. Prefer specific names like "AiAgent", "FieldCondition".
• `introspect_mutation` is expensive — fetches all root mutation fields and filters client-side (single large query). Prefer introspect_type on the specific input type when you already know the mutation name.
• `execute_graphql` returns GraphQL errors — check path and message; pass debug=true on the next call to surface the correlation_id.
• Endpoint confusion — introspection uses app.pipefy.com/graphql; real operations use api.pipefy.com/graphql. The MCP server handles this automatically; raw-API users must distinguish (see api-fallback).

See also

• docs/mcp/tools/introspection.md — MCP parameters, query/mutation mismatch hints on execute_graphql.
• skills/api-troubleshoot/pipefy-api-fallback/SKILL.md — Tier 3: direct HTTP fallback when MCP is unavailable.
• skills/pipes-and-cards/pipefy-pipes-and-cards/SKILL.md — most common dedicated tools (prefer over execute_graphql).