Loading ToolBench
Active state management for multi-agent coding: a shared, versioned project brain over MCP.
klypix-mcp shows moderate definition quality with significant gaps. Naming is generally verb-based and descriptive (brain_sync, list_canvases, read_canvas, recall, create_canvas, add_to_canvas, brain_note, brain_insights, brain_connect). Descriptions are present for all 10 tools and range from adequate to good (50-150 chars typical). However, critical issues emerge: (1) input schemas are visible for all tools, but parameter descriptions vary in quality and some are vague about expected formats/constraints; (2) output schemas are COMPLETELY ABSENT, no tool documents what it returns, forcing LLMs to guess at response structure; (3) error handling and recovery guidance are missing entirely; (4) no enums or format constraints for constrained parameters (e.g., brain_sync 'phase' has enum values visible in schema but no constraint description of what happens in each phase); (5) the brain_sync tool is problematic, its description mentions 'Codex Context Gateway coordination' with no explanation of what that means or when to use it vs other tools; (6) several tools (search_all_brains, brain_connect, brain_insights) lack clarity on expected output structure and pagination/limits. The tool set shows good naming discipline and verb-first conventions, but lacks the depth of parameter validation rules and output documentation that production-grade tools require.
Append a decision or cards (with optional connections) to an existing .klypix, preserving every existing item and position. The durable, cross-session memory a multi-agent run keeps writing to.
Analyze file changes and suggest brain connections: relate changed files to decisions/findings already on the brain.
Structural read of a brain.klypix: load-bearing hub cards, orphaned decisions, stale open questions, and area sizes. Use to orient before a planning task.
Record a note to the project brain (brain.klypix) with optional lifecycle marker and routing.
Automatic Codex Context Gateway coordination: synchronize active tasks, detect exact file overlaps, and deliver proactive alerts across MCP sessions.
Create a new .klypix canvas from cards + connections and return the board as a portable file artifact the human owns and any model can re-open.
NO OUTPUT SCHEMAS DOCUMENTED for any tool. Tools define inputs but provide zero guidance on return value structure, field names, types, or pagination. This forces LLMs to guess at response fields and breaks tool composition chains.
Enum values are described as literal strings in parameter descriptions ('', '?', '!', '+', '✓', '~') rather than formal JSON Schema enum constraints. LLMs cannot reliably parse unstructured enum lists from text.
Multiple search/recall tools (recall, search_all_brains) with overlapping functionality and no explicit disambiguation. 'search_all_brains' description lacks context for when to use it vs 'recall'. LLM will waste reasoning cycles deciding between them.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2025-06-18+ | v2 |
List every .klypix / .any canvas with card and connection counts.
Read one canvas as structured markdown — every card, the connection graph, [[wikilinks]], #tags — plus its images returned as file parts so a vision model SEES them.
Search card text, titles, and #tags across every canvas in the vault and return the matching cards — the shared blackboard a delegating agent reads before acting.
Search every project brain on this machine (cross-project capability).
brain_sync description uses jargon ('Codex Context Gateway coordination') without explanation. No clarity on why an agent would call this tool, what 'overlap detection' means operationally, or what 'proactive alerts' are. LLM cannot determine when to invoke it.
No error handling or recovery guidance documented for any tool. Missing actionable error messages for invalid inputs (e.g., 'canvas not found', 'invalid marker format', 'file glob matches nothing'). LLM has no guidance on how to recover from failures.
Pagination and result limits not documented. Tools like 'recall', 'list_canvases', and 'search_all_brains' may return unbounded lists. No limit parameter, no next_cursor field, no total count documented.
Parameter formats and constraints missing. 'color' parameter in create_canvas and add_to_canvas accepts any string with no format constraint (hex, rgb, name?). 'query' parameter accepts free-form strings with no guidance on syntax for tags or boolean operators.
Domain-specific terminology not explained. 'brain_insights' mentions 'load-bearing hub cards', 'orphaned decisions', 'stale open questions' without defining what these metrics mean or how they are calculated.
Tool composition chains broken. create_canvas and add_to_canvas accept 'from' and 'to' fields as 'string|number' (index, title, or id) with no clear output guidance. LLM cannot determine which identifier is returned for downstream reference in subsequent add_to_canvas calls.