Loading ToolBench
Self-growing procedural knowledge layer for AI agents — CLI and MCP server for managing evolving knowledge bases with critic feedback loops
Innate has 7 tools with explicit schemas and descriptions visible in TypeScript SDK. Naming is verb-based (recall, appraise, record, inspect, evolve, backup, daemon) and mostly clear. However, descriptions vary widely in quality, some are detailed (recall, appraise) while others are terse (inspect, daemon). Parameter descriptions are present but inconsistent; many lack format constraints, ranges, or actionable guidance. Output schemas are partially documented (RecallResult, InspectResult interfaces visible) but incomplete for all tools. Error handling guidance is absent, tools do not indicate what to do on failure. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (READ_ONLY vs WRITE).
Critic check: judge how much footing exists for a candidate answer in a given situation. Returns a Verdict with valence (affirm/caution/mixed/neutral), strength, tier, flagged points, and contributors — never an answer text.
Manage knowledge base backups to R2 (Cloudflare): run backup, show status, list backups, or prune old backups.
Manage the Innate daemon process: start, stop, restart, or check status. The daemon runs background evolution and embedding tasks.
Trigger knowledge base evolution: distill pending chunks into stable knowledge, curate active chunks (archive, promote, dedupe, decay), and rebuild embeddings.
Get comprehensive health and observability metrics for the knowledge base, including chunk composition, debt metrics, feedback loop statistics, daemon status, and performance trends.
Retrieve knowledge chunks from the procedural knowledge base using semantic similarity and context scoring. Returns ranked chunks with confidence scores, trigger descriptions, and optional spark suggestions.
Tool definitions inferred from SDK code, not explicit MCP tool registration visible in source. Cannot verify actual MCP server-side tool registration, schemas, or descriptions as served by the MCP protocol.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classification. Tools marked READ_ONLY or WRITE lack protocol-level hints to guide agent behavior.
Parameter descriptions lack format constraints, ranges, and actionable guidance. E.g., 'budget' (number) has no min/max; 'action' enums lack explanation of what each action does; 'source' parameter undocumented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Record the outcome of a traced operation (ok/fail/unknown) and optional feedback (up/down) to flow back into the knowledge base evolution loop.
Error handling and recovery guidance absent. Tools do not document what errors are possible, whether they are retryable, or what the agent should do next on failure.
Output schemas incomplete or missing for most tools. Only recall and inspect have TypeScript interfaces; appraise, record, evolve, backup, daemon lack documented return types.