Loading ToolBench

Reading the current corpus…

Cyrene MCP server · D · ToolBench

Cyrene

Open-source ReAct Agent framework oriented towards existing business systems, domain knowledge, and permission boundaries. Multi-language support with tool orchestration, sub-agent management, and document parsing capabilities.

2026-07-28+STREAMABLE_HTTP3 tools

57/100

Definition Quality66

Protocol Readiness65

Supportability38

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence · evidence: Streamable HTTP

Current-spec patterns detected

Summary

Cyrene Agent exposes 3 tools for sub-agent orchestration with complete JSON schemas and reasonable descriptions. However, descriptions are terse (avg 95 chars, below the 194-char baseline), parameter descriptions lack detail on constraints and formats, and error handling guidance is absent. Tool names follow verb_noun convention but lack clarity on when to use each (await vs get distinction is subtle). No output schema documentation visible. Schemas are present but sparse on validation rules.

Tools (3)

await_subagentsread onlysource verified68/100

Wait for sub-agent tasks to complete and get their results. All tasks share a single timeout deadline. Uncompleted tasks can auto-resume session.

cancel_subagentswritesource verified68/100

Cancel sub-agent tasks. Only tasks in the current run can be cancelled. Returns lists of cancelled tasks and already completed tasks.

get_subagentsread onlysource verified63/100

Get the status of sub-agent tasks without blocking. Returns task status, creation time, and results if completed.

Issues (5)

HIGHconstrained-input ↗

Parameter descriptions lack constraint details. 'timeout_seconds' has no min/max bounds; 'task_ids' array has no size limits. LLMs cannot infer valid ranges and may pass absurd values (timeout=999999, 10000 task IDs).

await_subagentscancel_subagentsget_subagents

HIGHresponse-shaper ↗

No output schema documentation. Tools return results but the response structure (fields, types, nested objects) is not documented. LLMs cannot plan downstream calls or extract required data reliably.

await_subagentscancel_subagentsget_subagents

HIGHrecovery-guide ↗

Error handling lacks recovery guidance. No indication of which errors are retryable, what the LLM should do on timeout, or how to handle partial failures (e.g., some tasks cancelled, others already complete).

await_subagentscancel_subagents

Recommendations

Add min/max bounds to numeric parameters: 'timeout_seconds (1 - 3600, default 120)', 'task_ids array (1 - 100 items)'. Include these in parameter descriptions, not just JSON Schema.
Document output schemas for all three tools. Specify: What fields does each return? Are results paginated? What does 'status' contain (enum values)? Include example response structures.
Expand tool descriptions to 100 - 200 chars. For get_subagents: 'Poll sub-agent task status without blocking. Returns task IDs, current status (pending|running|completed|failed), creation time, and results if completed. Use this to check progress; use await_subagents to block until completion.'
Add error recovery guidance. E.g., 'On timeout with on_timeout=RESUME_SESSION, the session auto-resumes when tasks complete. On CANCEL, remaining tasks are terminated and cannot be resumed.'
Clarify enum semantics in descriptions. For return_when: 'ALL = wait for all tasks to complete; ANY = return when first task completes (success or failure); FIRST_SUCCESS = return when first task succeeds (ignore failures).'
Add per-parameter validation rules. E.g., 'task_ids must be non-empty and contain valid UUIDs from the current run. Invalid IDs are silently ignored; if all IDs are invalid, returns empty result.'

Score history

Overall score trend

First recorded score · v2 rubric

57/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	57	2026-07-28+	v2