MCP server for TOTVS Fluig — read and write datasets, forms, global events and BPM process definitions over the API, without Fluig Studio.
Solid foundation with 25 well-named tools covering a complex domain (Fluig BPM/ECM). All tools have descriptions (avg 120 chars) and explicit input schemas with typed parameters. Naming follows verb_noun convention consistently (fluig_dataset_list, fluig_form_save, fluig_process_export_xml). However, output schemas are undocumented, responses are inferred from code, not declared. Parameter descriptions are present but often terse (e.g., 'Dataset name' vs 'The name of the dataset to query; case-sensitive'). Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors. Security is strong (credentials via env vars, confirm flags on destructive ops), but no tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is excellent, tools chain well (export_xml → import_xml, list → get → save). Missing: output schema documentation, richer parameter constraints (enums, ranges), and error recovery patterns.
DESTRUCTIVE. Delete a CUSTOM dataset from the server. Also how this server cleans up its own throwaway datasets.
Return the source code of a CUSTOM dataset.
List datasets on the server (id + type). Optional substring filter over id and description.
Run a dataset (form-backed or custom) and return its rows.
Create or update a CUSTOM dataset. The code runs on the server under Rhino, so it must be ES5 (no let/const, arrow functions or template literals).
Return a dataset's structure (columns and types) WITHOUT running it — useful to learn the schema before composing a query.
Output schemas are not documented. Tool descriptions state what is returned (e.g., 'return its rows'), but the response structure (field names, types, pagination) is not declared in the tool definition. LLMs cannot plan downstream calls or extract fields reliably.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Tools are marked write:true/false in code, but this metadata is not exposed via MCP tool annotations. Agents cannot infer safety without reading descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Run a SELECT against the Fluig database through a throwaway dataset. Read-only (SELECT/WITH only). Reaches any table, including event_proces, DEF_PROCES and the FDN_* metadata.
Return a form's customisation events (displayFields, validateForm, enableFields, ...) with their source.
Return the text content of one file of a form.
List the file names (HTML/JS/CSS/images) that make up a form.
Read a whole form: metadata, every file and every event. Do this before editing or publishing.
List forms (documentId, version, description, dataset). Optional substring filter.
Publish a form as a new (revertible) version. The API REPLACES the whole set, so send ALL files and events — read it with fluig_form_full first and change only what you need.
List the server's global events.
Create or update a global event.
Check authentication and session against the configured Fluig server.
Upload a process diagram (SVG). Fluig will render it in the Studio.
Extract events from a live process definition on the server (requires FluiggersWidget).
Extract events from a process definition XML (export it first with fluig_process_export_xml).
Export a process definition as ECM 3.0 XML (the format Fluig Studio uses).
Import a process definition from ECM 3.0 XML. Creates a new version.
List BPM processes (id, name, version, state).
List versions of a process.
Run a SELECT directly against the TOTVS RM database, when Fluig has a datasource for it. Read-only (SELECT/WITH only).
Query TOTVS RM through the stored-SQL bridge dataset: a registered statement code plus branch (CODCOLIGADA) and application (CODAPLICACAO). Read-only.
Error handling lacks recovery guidance. No tool documents what errors are retryable, what user action is needed, or what to try next. E.g., fluig_dataset_run could fail if dataset doesn't exist, but the error response doesn't suggest calling fluig_dataset_list first.
Parameter descriptions are terse and lack constraints. E.g., 'Dataset name' (3 words) vs 'The name of the dataset to query; case-sensitive; max 255 chars'. No enums for constrained values (e.g., versionOption in fluig_form_save accepts '0' or '2' but is documented as a free string).
Query tools (fluig_db_query, fluig_rm_db_query, fluig_rm_query) lack pagination and result limits. MAX_ROWS=200 is enforced in code but not declared in tool descriptions. If a query returns 200 rows, the agent doesn't know if there are more or if it's the full result set.