cartodb/agent-skills

carto-create-workflow

Builds, schedules, and operates analytics DAGs in CARTO Workflows — the no-code/low-code orchestration layer over the data warehouse.

View source
Original skill document

Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.

carto-create-workflow

CARTO Workflows is a visual DAG authoring app that compiles to warehouse SQL. Each workflow runs inside a connected warehouse — no CARTO compute is involved at execution time. This skill covers the full lifecycle: building the DAG (the bulk of this file), operating it via the CLI (CRUD, schedules), and cross-profile copy (dev → prod promotion, customer-segregated workspaces via carto workflows copy) — see the references below.

Workflows vs Builder — distinct apps. This skill targets Workflows (CARTO's DAG/orchestration app, /workflows/<id> in the URL). Builder is the separate map-authoring app (/builder/<id>). When this skill mentions "the Workflows canvas", "in Workflows", or "the DAG view", that is never Builder — they're different products that share an org and connections but nothing else. Workflow output (a result table with geometry) is typically visualised in Builder afterward; that's the only place the two apps connect.

For one-off ad-hoc SQL, use `carto-query-datawarehouse` — workflows are for repeatable, scheduled, multi-step DAGs.

Access-path routing. The interactive lifecycle maps to MCP tools on an OAuth-authenticated MCP session: create_workflow, update_workflow (method=update|share|unshare|publish|unpublish), validate_workflow (method=validate|verify|tosql), `readworkflows (method=list|get|list_mcp_tools|get_mcp_tool), runworkflow` (method=run|status|results), `scheduleworkflow (method=add|update|remove), and readworkflowcomponents (method=list|get — the component catalog). The phases below apply on either path — where a step names a carto workflows command, substitute the matching MCP tool (e.g. verify-remotevalidateworkflow method=verify`). Stay on the CLI for scripting/CI, bundle-schema introspection (`carto workflows schema` — no MCP equivalent), cross-profile copy (`carto workflows copy` — CLI-only), and usage-log auditing (`carto activity query` — CLI-only). Fall back to the CLI when the server isn't attached, or when the MCP session is **token**-authenticated — authoring/admin tools are hidden there; only `runworkflow and validate_workflow remain. Detection signals: [carto-basics/references/access-paths.md`](../carto-basics/references/access-paths.md).

Bundle structure, component schemas, input formats, and gotchas are all served live — by the CLI (carto workflows schema / components) or, on the MCP path, by read_workflow_components (method=list|get). Never hardcode or assume them. Live introspection is the source of truth.

Live introspection (use before reaching for any reference file). On MCP, read_workflow_components (method=list|get) covers the component rows; the bundle/DAG schema sections are CLI-only.

CommandWhat it serves
carto workflows schemaIndex of all bundle/DAG schema sections
carto workflows schema bundleTop-level bundle shape (id, title, connectionId, config, privacy, tags). privacy is a $ref — fetch its shape with carto workflows schema privacy. Minimal valid form: "privacy": { "privacy": "private" } (the inner string is not a bare "private"enums lists the allowed values for that inner field).
carto workflows schema configFull DAG config (schemaVersion, connectionProvider enum, nodes, edges, variables, viewport, useCache, executionSettings, schedule)
`carto workflows schema node[.source\.customsql]`Generic node shape (data.version, data.title vs data.label); .source adds the data.id == data.inputs[0].value invariant; .customsql the full customsql spec
carto workflows schema customsqlCopy-paste customsql node template (with version: "2.0.0")
carto workflows schema edge / schema handlesEdge shape; and the edge handle naming reference (sourceHandle/targetHandle by node type, operator, component — critical for valid edges)
carto workflows schema variableVariable (parameter) shape — { order, name, type, value, public }
carto workflows schema schedule / schema enumsDeclarative schedule metadata; and all valid enums (node types, providers, privacies, schedule frequencies)
`carto workflows components list \get <names> --connection <conn> --json`Component catalog; per-component inputs, outputs, notes. Add --input-formats to get for input-type format, examples, pitfalls. MCP: read_workflow_components.
carto workflows --helpFull command reference, including schedule-expression dialects per engine

References (only for what live introspection doesn't serve):

  • `references/providers/` — per-warehouse details (BigQuery, Snowflake, Databricks): identifier quoting, column casing, AT path.
  • `references/scheduling.md`add vs update semantics, bundle-level schedule warning, activity-log verification.
  • `references/mcp-and-api-publish.md` — publishing a workflow as an MCP tool or callable API endpoint (MCP: update_workflow method=publish): bundle requirements (native.mcptooloutput + scoped variables + draft descriptions), {{@var}} vs @var substitution syntax, Number → FLOAT64 LIMIT gotcha, post-publish verification.
  • `references/cross-profile-copy.md`workflows copy mechanics (CLI-only), connection mapping, --skip-source-validation, why copies are always new workflows.
  • `references/schedule-readd.md` — schedules don't transfer across workflows copy; how to re-add them, including dialect translation when source and destination engines differ.
`connectionProvider` must match the connection. config.connectionProvider (enum in schema enums) must match the connection's actual provider — mismatches generate the wrong SQL dialect and error at runtime. Look it up with carto connections list --search <name> --json (CLI) or explore_data (method=listconnections|getconnection) on the MCP path.

Development process

Follow these 6 phases in order for every workflow request. Do not skip or reorder them.

Phase 1 — Gather information

  1. Identify data sources. If the user named tables, note them. Otherwise discover what's available — CLI carto connections list / carto connections describe <connection> "<fqn>", or MCP explore_data (method=listconnections|listresources|describe).
  2. Clarify the goal. What transformation? What output? What filters/conditions?
  3. Determine the connection and note its provider (bigquery / snowflake / databricks) — you need it for the next step. CLI carto connections list or MCP explore_data method=list_connections.
  4. Read the provider reference.

<critical-rule id="read-provider-reference"> Before writing any node, you MUST open references/providers/<provider>.md (e.g. references/providers/bigquery.md) and read it end-to-end. This is non-negotiable.

<why>It contains identifier-quoting rules, column-casing behaviour, Analytics Toolbox path, schedule-expression dialect, and customsql $a/$b placeholder requirements that validate cannot catch. These only surface later as verify-remote failures or runtime SQL errors, and are the single most common cause of late-stage rework.</why>

<do-not>Do not skip this step because the next phases look concrete. Do not rely on memory of a previous run — provider files change.</do-not> </critical-rule>

  1. Fetch the component catalogcarto workflows components list --connection <connection> --json (CLI) or read_workflow_components method=list (MCP). Your only source of truth for component names.

Phase 2 — Design the approach

  1. Select components from the catalog you fetched.
  2. Fetch schemas for every component you plan to use. CLI carto workflows components get <name1>,<name2> --connection <connection> --json (or MCP read_workflow_components method=get) returns inputs, outputs, and notes. Read the notes array carefully — it contains gotchas.
  3. Fetch input type formats. CLI carto workflows components get <component1>,<component2> --connection <connection> --input-formats --json returns format, examples, and pitfalls for each input/output type. Pass component names (e.g. native.buffer), NOT input-type names.
  4. Design principles:
  • Preserve identifier and spatial columns throughout.
  • Prefer native components over `native.customsql`. This is not a soft preference. See Native-first rule.
  • H3/Quadbin columns work for visualization without geometry extraction.
  • Use standard names for visualization: geom, h3, quadbin.

Phase 3 — Present plan, surface gaps, confirm

Present the workflow plan (components, data flow, decisions). Then explicitly enumerate every gap before building:

  • Unresolved parameters — thresholds, radii, filter values, time windows, k for k-NN, aggregation columns, output table names, etc.
  • Analytical decisions left to the user — significance levels, distance metrics, join types, null-handling, dedup keys, CRS, H3/quadbin resolution.
  • Ambiguities in the request — anything where you had to guess intent.

For each gap, propose a sensible default with its rationale (e.g. "p-value threshold: suggest 0.05 — conventional significance level", "buffer distance: suggest 1000m — matches the city-block scale of the input"), and ask the user to confirm or override. Never silently pick a value for a user-facing analytical parameter. Wait for confirmation before building.

Phase 4 — Build the workflow

  1. Create the workflow file. Get the bundle/node/edge/variable shapes from carto workflows schema [section] (start with bundle, then node, node.source, node.customsql, edge, handles). For customsql nodes, copy the template from carto workflows schema customsql.

If you set the optional top-level privacy, it must be an object, not a string: "privacy": { "privacy": "private" } (the field name nests). Omit the field entirely if you don't need it — "privacy": "private" will fail validate.

Source nodes (type: "source") — treat ReadTable like any other component: fetch its spec with carto workflows components get ReadTable --connection <conn> --json to get the canonical inputs[*].title and inputs[*].description. (ReadTable is hidden from components list because it's grouped __internal, but get returns it normally.) Two source-only rules get cannot tell you, both from schema node.source:

  • The canvas display name lives in data.label, NOT data.title. Generic nodes use title; source nodes use label.
  • data.id and data.inputs[0].value must be the same FQN.

Canvas layout & naming — apply on every node, every workflow. None of this affects execution, but the user opens the DAG in Workflows and a sloppy canvas reads as low quality. The numbers are small and stable; just apply them.

  • Snap grid is 16 px. Every x and y must be % 16 == 0; off-grid values look subtly misaligned next to anything the user nudged.
  • Card widths by node type: source nodes 192 px (12 cells), generic components 64 px (4 cells). Heights fixed: every card is 80 px (5 cells) tall, with a 16 px label rendered below the body (the label lives in the gap to the next card, not inside the card).
  • Canonical inter-card gap (right edge → next left edge): 80 px (5 cells) tight/linear; 128 px (8 cells) at a fan-in (a join's left input, where an edge from another row needs room). The gap is the constant; left-edge Δx differs by card width — generic→generic linear = Δx 144 (9 cells), source→generic = Δx 272 (17 cells), generic→generic fan-in = Δx 192 (12 cells).
  • Canonical vertical gap (body bottom → next body top): 80 px (5 cells) = 16 px label + 64 px whitespace. So a stacked step is top-to-top Δy = 160 px (10 cells).
  • Layout. Source nodes stack at the leftmost column, same x, Δy = 144 px (9 cells). The main pipeline runs at the y-midline of the source rows (sources at y=80 and y=224 → pipeline at y=160), so joins receive both inputs symmetrically.
  • `data.title` and `data.label` are different fields — never duplicate. title = short instance-specific verb (≤ 15 chars) for what this node does ("Rank", "Join to score", "To H3"). label = the component's canonical type name on a fresh drop ("Join", "Create Column") — read from components get <name> --jsoncomponents[0].title. Source nodes only render data.label (a short alias for the table: "Candidates", "Score grid C").
  1. Run `validate` after every write to the file. It's offline, fast, and catches structural errors immediately — CLI carto workflows validate workflow.json --json, or MCP validate_workflow method=validate:
bash
   carto workflows validate workflow.json --json

Treat any save without a passing validate as broken — fix before continuing to the next node/edge.

`validate` is authoritative. If a component schema from components get disagrees with what validate accepts, trust validate and adjust the bundle to satisfy it. Do not "fix" the bundle to match the schema if it's already passing validation.

  1. Run `verify` at branch boundaries, not on every save. It hits the warehouse (slower, requires auth), so reserve it for whole sub-DAGs once their structure validates clean, and once at the end before presenting — CLI verify-remote, or MCP validate_workflow method=verify:
bash
   carto workflows verify-remote workflow.json --connection <connection-name> --json

verify is what catches column-type mismatches, missing tables, and AT resolution — things validate cannot see.

Reading the response. Walk deep.errors first — anything in there blocks runtime. deep.warnings is finer-grained: COMPONENT_INVALID items describe runtime concerns (an option the engine doesn't recognize, a column reference that may fail); SCHEMA_TRACE_SKIPPED and VERIFY_SKIPPED are diagnostic (the static checker punted on this node, not that it's broken). The top-level valid aggregates errors and warnings, so it can read false while the workflow still uploads successfully (deep.valid: true with empty deep.errors). Treat empty deep.errors as the upload gate; treat top-level valid: false as "look closer, but probably not blocking."

  1. Fix errors silently — don't expose implementation details to the user.
  2. Iterate until complete, with both validate and a final verify clean.

Phase 5 — Present result

Summarize what was built. Confirm validation success. Wait for user confirmation.

Phase 6 — Upload to CARTO

  1. Ask if the user wants to upload.
  2. Upload and provide the URL — CLI carto workflows create, or MCP create_workflow:
bash
   carto workflows create --file workflow.json --verify

The connection comes from connectionId inside the bundle — no --connection flag here.

  1. Confirm the upload didn't silently drop inputs. Immediately after create, read the saved workflow (carto workflows get <id> --json or read_workflows method=get) and diff config.nodes[*].data.inputs against the bundle you uploaded. The engine silently rejects values at save-time when an input fails validation (most commonly a Selection input fed a display label instead of a wire value — see "Display labels vs wire values" in Fetching component & input information). When this happens, validate, verify (when deep.valid: true with deep.warnings only), and create all report success; Workflows renders the node with a red error indicator on first open. Any value present locally but missing on the server was silently rejected — fix the source bundle and re-upload (don't try to edit on the server).
  2. Do NOT auto-execute unless explicitly requested.

Native-first rule

native.customsql is the last tool to reach for, not the first. Before writing a customsql node, attempt the native chain. Fall back to customsql only if at least one of these is true:

  • The native chain would require more than ~4-5 nodes to express the same logic.
  • A specific operation has no native equivalent at all (verified via carto workflows components list).
  • The expression genuinely needs raw warehouse SQL (e.g., ST_UNION_AGG, LOGICAL_OR, ML.PREDICT, last-N windowing).

Common operations and their native equivalents — try these first:

If you'd write SQL like…Use natives
WHERE x = … / multi-condition filternative.where (predicate), native.wheresimplified (form-based filter UI), native.spatialfilter (geometry-based match/unmatch split), native.select (column projection)
SELECT a, b, c FROM t (multi-column projection / rename / multi-expression)native.select (one node, free-form SELECT body)
SELECT ..., expr AS c FROM t (add one computed column)native.selectexpression (one column + one expression per node)
GROUP BY k, SUM(x), AVG(y), COUNT(*) (single key)native.groupbygroupby input is a single Column, not multi-column. For multi-key grouping use native.customsql.
JOIN ... ON a.k = b.k (any join type)native.joinv2
JOIN ... ON ST_INTERSECTS / ST_CONTAINS / ST_WITHINnative.spatialjoin
MIN(ST_DISTANCE(a.geom, b.geom)) across two tablesnative.distance (augments the main table in place with nearest_id + nearest_distance — rename them per source if you chain two native.distance nodes for two reference tables)
ST_BUFFER(geom, d)native.buffer
H3 binning / boundary / center / polyfillnative.h3frompoint, native.h3boundary (output geometry column is named <h3col>_geo, e.g. index_geonot geom), native.h3center, native.h3polyfill
ORDER BY ... LIMIT nnative.orderby + native.limit
z-score / standardizationnative.normalize
weighted composite scorenative.spatialcompositeunsupervised (weighted/PCA), native.spatialcompositesupervised (target-driven)
Getis-Ord Gi*, GWR, isolinesnative.getisord, native.gwr, native.isolines
Save final node to a tablenative.saveastable

Signals you're reaching for customsql too early — stop and look for a native chain instead:

  • The customsql is just a WHERE clause, a single JOIN, a GROUP BY with one or two aggregates, or a column projection.
  • It wraps a single warehouse function (ST_BUFFER, H3_FROMGEOGPOINT, etc.) for which a dedicated native exists.
  • Its only purpose is to project/rename/re-cast columns — use native.select (free-form SELECT body, one node) for multiple columns; native.selectexpression is for adding a single computed column.
  • You're chaining customsql outputs through more customsql nodes — chain natives instead.

When customsql is genuinely the right call, the per-warehouse SQL-dialect footguns live in the matching references/providers/*.md (BigQuery backticks, Snowflake casing, Databricks identifiers).

Customsql schema-trace cascade — misleading "Table X used but not provided" errors. If verify-remote reports Table "X" used in statement but not provided for a customsql node whose sourcea / sourceb are clearly wired, the root cause is almost always an upstream node with a failing schema trace — even a warning-level COMPONENT_INVALID on the upstream node is enough to break the cascade and surface on the downstream customsql. Fix the upstream node first; the downstream error will resolve. Do not spend cycles rewriting the SQL, swapping backticks, or experimenting with aliases until upstream is clean.


Fetching component & input information

Do not rely on memorized component schemas or input formats. Always fetch live — carto workflows components list | get <names> [--input-formats] --connection <conn> --json (CLI) or read_workflow_components (method=list|get) on the MCP path (see the introspection table above).

What to look for in the response:

  • Component `notes` — gotcha strings: non-obvious behavior, deprecated status, output column naming.
  • Input `format` — prose describing the expected value shape.
  • Input `examples` — concrete JSON snippets showing correct usage.
  • Input `pitfalls` — common mistakes, evaluation order, format quirks.
  • Component `version` — copy verbatim into the authored node's data.version (string). Generic nodes without it are flagged OUTDATED in Workflows.
  • Input `options` (Selection / Enum) — the engine matches values exactly. Copy each option string verbatim — preserve case, never paraphrase or Title-Case (e.g. spatialjoin's jointype accepts "inner", not "Inner").
  • Display labels vs wire values. Some components (e.g. native.isolines.mode) carry a separate optionsText field for human-readable labels, and components get --json may surface those display labels under the options key. If verify-remote rejects your "verbatim" value with Valid options: <list> listing the opposite case, the CLI fed you display labels — find the true wire value by cross-referencing a known-good workflow with carto workflows get <id> --json. Lowercased / snakecased forms (`walk`, `publictransport`) are common for v2 components.
  • `verify-remote` may run multiple component versions simultaneously. If errors and warnings list contradictory "valid options" for the same input (e.g. one accepts "Walk", the other accepts "walk"), the engine likely ran both v1 and v2 validators against your node. Trust the version you copied from components get, and confirm by uploading a single-node test workflow.

For values that may evolve over time (component versions, bundle/config defaults, enum option lists), treat the CLI's components get / schema output as the single source of truth — never hardcode values in your own templates. Specifically:

  • `config.schemaVersion` — read the current default from carto workflows schema config --jsonproperties.schemaVersion.default. Today it's "1.0.0" (string), but resolve at author time so future bumps don't require a skill update.

Provider-specific notes

Different warehouses have different SQL dialects, table-naming conventions, and column-casing rules — always check the matching provider guide (bigquery, snowflake, databricks). Input-type formats (Table, Column, ColumnsForJoin, SelectColumnAggregation, etc.) and per-component gotchas (including the "AT components need verify, not validate" rule) come from live introspection — see Fetching component & input information.


Operating a workflow (after it's built)

Once a workflow exists in CARTO, both paths expose CRUD, run, and schedule management. MCP: read_workflows (list|get), update_workflow (update|share|unshare|publish|unpublish), run_workflow (run|status|results), schedule_workflow (add|update|remove), delete kind=workflow. CLI quick reference:

bash
carto workflows list --json                                   # list / inspect
carto workflows get <id>
carto workflows update <id> --file workflow.json              # update (replaces whole DAG)
carto workflows schedule add <id> --expression "every day 08:00"
carto workflows schedule remove <id>

Always-on guidance:

  • Workflows run on the connection's warehouse. A workflow with a BigQuery connection cannot use Snowflake-specific SQL.
  • Schedule expression syntax depends on the engine — natural-language for BQ/CARTO DW ("every day 08:00"), cron for Snowflake/Postgres ("0 8 * * *"), Quartz cron for Databricks ("0 0 8 * * ?"). See `references/scheduling.md`. Picking the wrong dialect fails at schedule-add time.
  • Copying a workflow across profiles (dev → prod, customer-segregated workspaces) is CLI-only — see `references/cross-profile-copy.md`. Schedules don't transfer — see `references/schedule-readd.md`.
  • Deleting a workflow doesn't delete its outputs. Tables/views the workflow created in the warehouse persist; clean them up with carto sql job (or MCP execute_query) if needed.
  • `update` replaces the whole DAG. There's no per-node patch. Always get first, edit, then update.
  • Workflow execution status lives in the activity log (WorkflowRun, WorkflowExecutionComplete event types). For health monitoring of scheduled workflows, query that log via `carto-query-datawarehouse` — see references/activity-queries.md in that skill.
from this repository

More skills

All skills