Loading ToolBench
Local MCP server for capturing screenshots, recordings, and reviewable evidence from long-running agent tasks.
Server has 15 tools with complete input schemas and descriptions. Naming follows verb_noun convention (start_session, capture_checkpoint, list_sessions). Descriptions are present and substantive (avg ~100 chars), exceeding the 20-char minimum. However, descriptions lack actionable context for LLM selection (e.g., when to use capture_checkpoint vs capture_screenshot). Parameter descriptions are generic ('Path to the session directory' repeated 15 times). Output schemas are defined as dataclasses but not documented in tool descriptions. Error handling is absent, no guidance on retryability, user-fixable errors, or recovery paths. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Missing dependency hints (e.g., 'Call start_session before capture_checkpoint'). Composition is sound, each tool has one responsibility, but chaining is implicit rather than explicit.
Attach a human-readable note to an existing artifact.
Capture a screenshot checkpoint for an existing session.
Capture a raw screenshot artifact for an existing session.
Compare two artifacts and return a detailed analysis of their differences.
Mark a session as completed and refresh its summary.
Load a session by session id or session directory path.
Return the newest evidence session.
No error handling guidance. Tools lack recovery hints (e.g., 'Session not found. Call list_sessions() to see available sessions'). LLMs cannot self-correct on failures.
Descriptions lack actionable context for tool selection. 'Capture a screenshot checkpoint' does not explain when to use capture_checkpoint vs capture_screenshot vs start_recording. LLMs cannot distinguish.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot infer which tools are safe to retry or have side effects. start_session, end_session, capture_* are destructive but unmarked.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
List captured artifacts for a session.
List evidence sessions, newest first. Optionally filter by status or limit the number returned.
Attach OCR text to an existing artifact, either manually or via the configured OCR backend.
Show the active recording state for a session.
Create a redacted screenshot artifact from an existing screenshot.
Start a short recording for an existing session.
Create a new evidence session for a long-running task.
Stop the active recording for a session and save it as an artifact.
Parameter descriptions are generic and repeated ('Path to the session directory' appears 15 times). No format hints, constraints, or examples. LLMs cannot validate inputs before calling.
Output schemas defined as dataclasses but not documented in tool descriptions. LLMs cannot plan downstream calls without knowing what fields are returned.