Loading ToolBench
Local-only observability + control plane for every AI coding agent on your machine. TUI live tail + browser dashboard on localhost. Unified timeline across Claude Code, Codex, Gemini CLI, Cursor, Hermes, OpenClaw — token + cost accounting, compaction + anomaly detection, SVG call graphs, diff attribution, agent-aware replay, MCP server mode, OpenTelemetry exporter. No cloud, no telemetry, no sign-in.
Five tools with clear verb-noun naming (list_, get_, search_) and solid descriptions (100-200 chars). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented in the source code, critical for LLM planning. Parameter descriptions lack constraint details (ranges, formats, enums). No error handling guidance visible. Tools are read-only (good for safety) but lack idempotent/readonly annotations. Descriptions are functional but could be more LLM-optimized per pattern:tool-description baseline (194 chars avg, 10-1024 range).
Return total cost (USD), token counts broken down by input / cache read / cache create / output, and turn count for a given session.
Return the raw events for a given session ID. JSONL for file-based agents (Claude/Codex/Gemini/OpenClaw); Hermes messages are serialized as one JSON object per line. Use after list_recent_sessions to drill into a session.
Aggregate tool invocation counts, total duration, and error counts. If sessionId is given, stats are scoped to that session; otherwise scoped to the N most recently active sessions across all agents (default 50).
List the most recent local agent sessions across Claude Code, Codex, Gemini, OpenClaw, and Hermes, newest first. Use to find a session to inspect.
Substring search across all local agent session files. Returns matching sessions with the first few matching lines. Covers Claude, Codex, Gemini, OpenClaw, and Hermes.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., what fields does list_recent_sessions return? Is there pagination?). Critical for tool composition.
Parameter constraints missing. 'limit' params lack min/max bounds (e.g., get_session_events maxBytes 1024-10000000 is stated in description but not in schema constraints). 'query' in search_sessions has no length/format constraints. Per pattern:constrained-input, enums and ranges must be machine-parseable.
No tool annotations (readOnlyHint, idempotentHint). All five tools are read-only and safe to retry, but LLMs cannot infer this without explicit annotations. Per current MCP spec (2026-07-28), tool annotations enable safer agent planning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
No error handling guidance. If a sessionId is invalid or not found, what should the LLM do? No recovery hints (e.g., 'Try list_recent_sessions first'). Per pattern:recovery-guide, errors must guide next steps.
Pagination not documented. list_recent_sessions and search_sessions return lists but no mention of total count, next_cursor, or whether results are capped. Per pattern:paginated-result, large result sets must support pagination.