Loading ToolBench
Git-native context layer for AI-assisted development. MCP server that manages project and feature context documents for AI-assisted development workflows.
Four tools with complete Zod schemas and descriptions. Naming follows verb_noun pattern (get_, list_, update_, export_). Descriptions are adequate (50-150 chars) but lack LLM-optimized guidance on WHEN to call each tool or dependencies between them. All parameters have types and descriptions. Output schemas are not documented in the source, only input schemas visible. Error handling exists but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Export context formatted for a specific LLM (claude, openai, gemini, cursor). Use when switching tools or sharing context.
Get the anchored context for this project. Always call this at the start of a task.
List all feature context documents available in this project.
Append a new decision to a feature context document after a decision is made.
Output schemas not documented. LLMs cannot infer what fields get_context, list_features, update_feature, and export_context return. This forces agents to guess structure and risks failed downstream tool chaining.
Tool annotations missing. update_feature is destructive (WRITE risk) but has no destructiveHint. get_context and list_features are read-only but lack readOnlyHint. Agents cannot distinguish safe from risky operations.
Descriptions lack WHEN-to-call guidance. 'Get the anchored context' does not explain when to call get_context vs export_context, or whether to call list_features first. LLMs must infer intent from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Error handling lacks recovery guidance. Code returns error strings ('Feature not found') but does not suggest next steps (e.g., 'Call list_features to see available features'). Agents cannot self-correct.
list_features lacks pagination parameters. If .context/features/ grows large, returning all features at once wastes tokens and risks context window exhaustion. No limit or offset parameters visible.