Reliable tool-calling agents on local SLMs. Hybrid REWOO/ReAct architecture, MCP support, one-command export. No API keys, no cloud.
OnsetLab exposes a single MCP tool 'DateTime' with significant definition quality gaps. While the tool has a description and enumerated operation parameter, critical issues emerge: (1) The tool name 'DateTime' is generic and does not follow verb_noun convention (should be 'get_datetime' or similar). (2) The description is present but terse (32 characters: 'Get current date/time, day of week for a date, add days to a date, or calculate days between dates'). While within bounds, this description conflates four distinct operations, suggesting the tool violates single-responsibility principle, each operation (now, day_of_week, add_days, difference) should arguably be its own tool or the description should be expanded to ~150+ chars to provide clear guidance on when to use each. (3) Parameter descriptions are minimal: 'Date (accepts: 2000-01-01, January 1 2000, Jan 1 2000, etc.)' provides format guidance but lacks detail on timezone behavior, parsing edge cases, or whether future dates are accepted. (4) The 'operation' parameter uses an enum (correct pattern), but other parameters (date, date2, days, timezone) lack type definitions in the visible schema snippet, only descriptions are shown. (5) No output schema is documented; the agent cannot reason about what fields the response contains. (6) The tool lacks error handling guidance: what happens if an invalid date is passed? If date2 is omitted for 'difference' operation? (7) Parameter 'timezone' defaults to 'local' but is listed as optional, no guidance on default behavior or supported timezone formats. The source code snippet is truncated and does not show complete tool registration, limiting full assessment.
Get current date/time, day of week for a date, add days to a date, or calculate days between dates.
Tool name 'DateTime' does not follow verb_noun convention. Generic nouns fail to signal intent to LLMs. Should be 'get_datetime', 'calculate_date_difference', or split into separate tools per operation.
Single tool 'DateTime' bundles four distinct operations (now, day_of_week, add_days, difference) under one operation enum. Violates single-responsibility principle. Each operation should be a separate tool (get_current_datetime, get_day_of_week, add_days_to_date, calculate_date_difference) so LLMs can compose and reason about them independently.
Parameter descriptions lack actionable detail. Example: 'Timezone (e.g., UTC). Default: local.' does not specify (a) supported format (IANA names? Offset notation?), (b) whether all operations respect timezone or only 'now', (c) what 'local' means in a containerized/serverless environment. Descriptions must answer WHAT, WHEN, and HOW.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 41 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
Output schema is not documented. No response structure shown for each operation. LLMs cannot plan downstream use or validate output format. Must document return type for each operation (e.g., 'now' returns {datetime: string, timestamp: integer}, 'difference' returns {days: integer}).
No error handling or recovery guidance. What happens if date parsing fails? If 'days' is not provided for 'add_days' operation? If timezone is invalid? Error responses must tell the agent what to try next (e.g., 'Invalid date format. Try YYYY-MM-DD or Month D YYYY').
Parameter 'date' description includes examples ('2000-01-01, January 1 2000, Jan 1 2000') but lacks formal constraint. LLMs often reuse examples literally or misinterpret boundaries. Use an enum, regex pattern, or explicit format string (ISO 8601, natural language patterns supported) in the schema, not narrative description.
Visible source code truncates at 'params[param_name] = {"type'. Full schema structure not visible. Cannot verify (a) whether all parameters have 'type' fields, (b) whether 'required' array is set correctly, (c) whether descriptions are complete in schema or only narrative. Truncation prevents full assessment; assume worst case (missing types/descriptions in schema).