Loading ToolBench
Your whole desktop context — everything you've seen, read, typed, and heard — as a single MCP your agent can use to understand you and how you work. Open source, local-first, on-device.
Three tools with complete input schemas and strong, LLM-optimized descriptions (150 - 250 chars each). All parameters are typed and described. However, output schemas are not documented in the visible code, only inferred from implementation. No error handling guidance, no parameter constraints (enums), and no explicit mention of pagination or result limits. Tool names are action-oriented (recall, catch_up, profile) but lack verb prefixes. Descriptions excel at explaining WHEN to use each tool and what context they provide, aligning well with agent decision-making. Missing: output field documentation, error recovery paths, and validation rules.
Get the user's recent activity (newest first) with no query needed — a snapshot of what they're working on and thinking about right now. Call it to orient yourself before responding so your help aligns with their current focus and style. Use `window` to pick the range.
Understand WHO this user is so you can tailor your help to them — what they're building, recurring people/projects/tools, how they think and write, their taste. Call it before giving generic advice on open-ended, creative, or work-related tasks so your answer fits them. Pass `topic` to focus (e.g. their taste in a specific area).
Find specific moments from the user's own recent activity across all their apps — code, writing, reading, decisions, conversations. Use it to recall a specific thing they reference, learn how they did something before, or find what was decided. Reach for it proactively, on your own judgment, whenever grounding your answer in what the user has actually done would make it more accurate or specific — don't wait to be asked. Filter by time/app/source to narrow.
Output schemas not documented. Tool descriptions explain what is returned in prose, but no formal JSON Schema for response fields. LLMs cannot plan downstream operations or extract specific fields without explicit schema.
No error handling guidance. Descriptions do not explain failure modes (e.g., 'no results found', 'query too broad', 'access denied') or recovery steps. Agents cannot self-correct on failures.
Missing parameter constraints. 'window' and 'since'/'until' accept free-form strings with no enum or pattern validation. LLMs may pass invalid values like 'last_week' instead of 'week'.
Tool names lack action verbs. 'recall', 'catch_up', 'profile' are nouns/gerunds, not verb_noun patterns (e.g., 'search_activity', 'get_recent_activity', 'get_user_profile'). Reduces clarity for LLM tool selection.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
No pagination or result limit documentation. 'recall' defaults to k=5 but no max stated; 'catch_up' defaults to limit=12 but no bounds. Large result sets could exhaust context windows.