Loading ToolBench
Local-first workflow compiler for AI agents that records privacy-minimized observations, learns from repeated work patterns, and compiles evidence-backed workflows into portable Agent Skills.
EvoPilot defines 18 tools with complete input schemas and descriptions. Naming follows verb_noun convention (observe, remember, analyze, compile, install). Descriptions are substantive (avg ~120 chars) and explain purpose clearly. However, parameter descriptions lack detail, most omit format constraints, ranges, or usage guidance. Output schemas are undocumented; LLMs cannot predict response structure. Error handling is minimal; no recovery guidance or actionable error messages visible. Security patterns (approval IDs, one-time tokens) are well-designed but not reflected in parameter descriptions. Composition is sound, tools are single-responsibility and chainable via fingerprints/IDs.
Detect repeated two-to-four-step workflows and measure their outcomes.
After explicit human confirmation, authorize the exact action ID blocked by the safety gate for one use within ten minutes.
Compile an evidence-backed workflow into a portable Open Agent Skills bundle for validation and review.
Retrieve concise relevant memories and learned workflows for the current scope.
Explicitly correct a memory while preserving its history.
Check the local runtime, required plugin files, and database health without returning stored memory content.
Output schemas undocumented. No tool describes what fields it returns, forcing LLMs to guess response structure and plan downstream calls blindly.
Parameter descriptions lack format constraints and usage guidance. E.g., 'scope' parameter has no explanation of valid values, format, or examples. 'fingerprint' is never defined, is it a hash, UUID, or opaque token?
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. E.g., evopilot_install_skill could fail if approval_id is invalid or expired, no recovery path documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Delete one memory by exact key while retaining a deletion audit event.
Install one reviewed Skill bundle after explicit user confirmation and exact one-time approval.
Record a privacy-minimized work observation and outcome for later workflow analysis.
Record one privacy-minimized behavior-cloning episode as structured steps, validation, and outcome.
Validate a generated Skill and return its evidence plus the exact one-time approval ID required for installation.
Promote a qualified behavior workflow into a short reviewed policy card for token-capped runtime retrieval.
Store a non-sensitive explicit or inferred preference. Inferences never silently replace conflicting values.
Retire a behavior policy so it is no longer returned in runtime context.
Classify a proposed action. Unknown or dangerous actions require a person.
Return a deterministic, token-capped behavior policy context for a task type without replaying episodes.
Structurally validate and score a compiled Skill bundle without executing or installing it.
Produce a concise evidence-based learning report for the last 1-90 days.
Destructive operations (evopilot_forget, evopilot_retire_policy, evopilot_install_skill) lack confirmation or dry-run patterns. No mention of audit trails or undo mechanisms.
STDIO transport only. Server is not remotely accessible and cannot integrate with hosted MCP clients. Limits deployment to local-only scenarios.