Harvest MCP has 18 tools with generally consistent naming and clear descriptions. Most tools follow verb_noun convention (get_, list_, create_, analyze_). However, there are significant gaps: (1) Input schemas are present but vary in completeness, some parameters lack type constraints and validation. (2) Output schemas are not documented in the source; responses are inferred from API mapping, not explicitly shaped for LLM consumption. (3) Parameter descriptions are minimal (1-2 sentences) and lack format constraints, ranges, or examples. (4) Complex tools like analyze_profitability and analyze_budget_performance have deeply nested 'object' and 'array' parameters with insufficient sub-schema documentation. (5) Error handling is not visible in the tool definitions. (6) The tool descriptions are adequate (50-150 chars) but could be more explicit about when to use each tool vs. similar ones. Overall, the server shows solid foundational structure with clear verb-noun naming and basic descriptions, but falls short of production-grade documentation for complex analytics tools.
Aggregate time entries by specified dimensions (client, project, user, task, date, week, or month)
Analyze employee performance based on budget adherence per project
Analyze profitability metrics across your Harvest data (projects, clients, users, tasks, cost centers, or custom)
Analyze utilization metrics (capacity, billable/non-billable ratios) for users, teams, or the whole account
Get schema documentation for Harvest data types, fields, enums, and relationships
Get billable and cost rates for a specific user (applies to project assignments or global defaults)
Create a new time entry
Output schemas not documented. Complex analysis tools (analyze_profitability, analyze_utilization, aggregate_time, analyze_budget_performance) have no visible return schema documentation. LLMs cannot plan downstream tool calls or validate response structure without seeing output field definitions.
Complex parameters lack sub-schema documentation. Tools like analyze_profitability, analyze_utilization, and analyze_budget_performance accept 'date_range' (object) and 'group_by' (array) parameters with no nested field definitions visible. LLMs cannot infer what keys or values these objects/arrays require.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 66 | 2026-07-28+ | v2 |
| 2026-03-09 | C | 60 | - | v1 |
Get information about the authenticated Harvest company/account
Get information about the currently authenticated Harvest user
Get a specific project by ID
Get a specific time entry by ID
List all clients in the Harvest account with optional filters
List all invoices in the Harvest account with optional filters
List all projects in the Harvest account with optional filters
List all tasks in the Harvest account with optional filters
List time entries with optional filters (user, client, project, date range, etc.)
List all users in the Harvest account with optional filters
Resolve entity names/identifiers to their IDs (clients, projects, users, tasks, teams)
Parameter descriptions lack format constraints and validation rules. E.g., 'from' (string) and 'to' (string) parameters in harvest_list_time_entries say they are dates in 'YYYY-MM-DD' format, but LLMs cannot validate this without explicit pattern/regex constraints in the schema or clear ranges in descriptions. Same issue for 'updated_since' (ISO 8601 format not enforced).
harvest_create_time_entry is under-specified. Description says 'Create a new time entry' but does not document what fields are required vs. optional, what happens on duplicate (is it idempotent?), whether it starts a running timer, or what the response contains. LLMs cannot safely call this without understanding side effects.
No error handling guidance visible. Tool definitions do not indicate what errors are possible (e.g., 'user not found', 'invalid date range', 'permission denied') or how the LLM should recover. This violates the recovery-guide pattern.
Pagination semantics not explicit. Tools like harvest_list_users and harvest_list_time_entries support 'auto_paginate' flag but do not document whether it fetches all pages or caps at 10, whether total_count is returned, or how large result sets are handled. LLMs may assume unlimited results and exhaust context.
Numeric parameter bounds not specified. Tools accept 'page', 'per_page', 'limit' parameters but do not document min/max values. LLMs could pass per_page=10000 (causing API errors or timeouts) or page=999999 without bounds-checking.
Missing tool composition context. No clear guidance on which tools call which. E.g., does resolve_entities return a user_id that harvest_list_time_entries accepts? Do list tools return sufficient detail for get_ tools to work? Broken chains force discovery detours.