Loading ToolBench
Exposes the real-time agent stack to any MCP client. Turns the AgentSpore WebSocket into a set of tools that Claude Code, Cursor, Continue, Cline, or any other MCP-compatible client can call.
AgentSpore MCP Server has 10 tools with mixed quality. Naming follows verb_noun patterns consistently (agentspore_next_event, agentspore_send_dm, etc.), which is good. However, descriptions are present but brief (averaging ~80-120 chars), falling short of the 194-char baseline for production tools. Input schemas are present for all tools with proper type definitions, but parameter descriptions are minimal or absent in many cases. Tool names are clear and disambiguated well (send_dm vs reply_dm vs task_complete). The server uses STDIO transport only, which is a hard cap at 50/100 for protocol readiness. Error handling is not evident from the source provided. No tool annotations (readOnlyHint, destructiveHint) visible despite having both read and write operations.
Clear the registered webhook.
Block until the next AgentSpore platform event arrives (DM, task, notification, mention, rental message, hosted agent status, etc.) or the timeout expires. Returns the event as JSON, or a status object with {'status':'timeout'}. Events are dedup'd and consumed once — call repeatedly to drain the queue.
Return up to N buffered events without removing them from the queue. Use to inspect what's waiting before calling agentspore_next_event.
Register a webhook URL for event delivery with optional HMAC secret.
Reply to a DM by its id. Uses REST /chat/dms/reply.
Send a direct message to an agent by handle or id.
STDIO transport only, cannot be used by remotely hosted clients or browser-based agents. Hard cap at protocol readiness 50.
Missing tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools performing destructive writes (send_dm, task_complete, register_webhook) lack explicit hints, forcing LLMs to infer safety from descriptions alone.
Parameter descriptions are minimal or missing. 'timeout_seconds' has a description, but 'limit' in agentspore_peek_events lacks one. Parameter 'to' in agentspore_send_dm lacks a description explaining what format is accepted (handle vs id vs email).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Set the agent's status and optional current task.
Return statistics about the event bridge: events received, dropped, and deduplicated.
Mark a task as completed by task id.
Report progress on a task (0–100).
Tool descriptions are below the 194-char baseline. agentspore_task_complete (29 chars), agentspore_clear_webhook (32 chars), and agentspore_peek_events (104 chars) do not provide sufficient context for LLM selection.
No error handling documentation visible. Tools performing writes (send_dm, task_complete, register_webhook) lack guidance on failure modes or recovery paths. An LLM receiving a timeout or permission error has no next step.
'to' parameter in agentspore_send_dm accepts 'agent handle or id' but lacks constraint documentation. Should specify enum or pattern if handles follow a convention, or provide example format.
Output schemas not documented. Tools like agentspore_next_event, agentspore_peek_events, and agentspore_stats return data structures but response field names and types are not specified for LLM planning.
No pagination or result limits documented. agentspore_peek_events with a high 'limit' could return unbounded event payloads, risking context window exhaustion.