Loading ToolBench

Reading the current corpus…

tabulus MCP server · C · ToolBench

tabulus

Postgres MCP server — agent-first database workbench. Exposes read-only SQL tools for AI agents to query PostgreSQL databases safely with PII redaction.

2026-07-28+STDIO5 tools

61/100

Definition Quality77

Protocol Readiness50

Supportability40

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

Tabulus demonstrates solid fundamentals: all 5 tools have clear verb-noun names (list_, describe_, sample_, safe_, explain), explicit input schemas with type definitions, and substantive descriptions (avg 120 chars). Parameters are typed and mostly described. However, output schemas are undocumented, LLMs cannot predict response structure. Error handling is basic (generic exception catches). No parameter constraints (enums, ranges, patterns) despite opportunities (e.g., safe_select should reject non-SELECT). Missing descriptions for some parameters (e.g., 'table' in sample_rows lacks detail on schema qualification). Security is strong (read-only enforcement, redaction support, audit logging), but composition could be tighter, describe_schema and sample_rows overlap in intent.

Tools (5)

describe_schemaread onlysource verified82/100

Describe a table: columns, types, primary key, foreign keys, indexes, and a small sample of rows. The single most useful tool for an agent trying to understand the data model.

explainread onlysource verified72/100

Return the query plan for a SELECT statement (EXPLAIN FORMAT JSON).

list_tablesread onlysource verified82/100

List all tables in the database with row count estimates and sizes.

safe_selectread onlysource verified77/100

Run a read-only SELECT query. INSERT/UPDATE/DELETE/DDL are rejected. Results are capped at the server's max_rows setting (default 100).

sample_rowsread onlysource verified73/100

Return a random sample of rows from a table.

Issues (5)

HIGHtool ↗

Output schemas undocumented. LLMs cannot predict response structure (e.g., does list_tables return {tables: [...]} or [{name, rows, size}, ...]?). Forces agents to guess field names and types.

list_tablesdescribe_schemasample_rowssafe_selectexplain

MEDIUMtool-description ↗

Parameter 'table' in sample_rows and describe_schema lacks description detail. Does it accept 'schema.table' or just 'table'? Ambiguity invites errors.

sample_rowsdescribe_schema

MEDIUMconstrained-input ↗

No input constraints (enums, ranges, patterns). 'limit' in sample_rows has no min/max; 'sql' in safe_select has no format hint. Agents may pass invalid values.

sample_rowssafe_selectexplain

MEDIUMrecovery-guide ↗

Error handling returns generic exception names ('UnsafeSQLError', 'Error: Exception: ...') without recovery guidance. LLM cannot determine if error is retryable or user-fixable.

Recommendations

Document output schemas for all tools. Example: 'Returns {tables: [{name: string, row_count: int, size_bytes: int}, ...]}' in list_tables description or a separate outputSchema field.
Add parameter descriptions: 'table' should clarify 'Table name, optionally schema-qualified (e.g., public.users or just users).' Expand 'limit' to '(default 10, max 100).'
Add input constraints: 'limit' → {type: integer, minimum: 1, maximum: 100}; 'sql' → {type: string, pattern: '^(SELECT|WITH|EXPLAIN|SHOW)'}.
Enhance error messages with recovery hints. Instead of 'Rejected by safety policy: ...', return 'SQL rejected: only SELECT/WITH/EXPLAIN/SHOW allowed. Rewrite without INSERT/UPDATE/DELETE and retry.'
Clarify describe_schema vs sample_rows in descriptions. E.g., describe_schema: 'Returns schema metadata (columns, types, keys, indexes) plus a small sample.' sample_rows: 'Returns random data rows only, use describe_schema for structure.'

Score history

Overall score trend

First recorded score · v2 rubric

61/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	C	61	2026-07-28+	v2