Loading ToolBench
A self-contained MCP server for AI agent task scheduling and file watching
39 tools with consistent verb_noun naming (agent_*, loop_*, spec_*, etc.). All have descriptions (10-150 chars, mostly adequate). Input schemas present with typed parameters and descriptions. However, output schemas are undocumented, LLMs cannot predict response structure. Parameter descriptions lack format constraints (e.g., RFC3339 format for timestamps, path validation rules). No enum constraints on platform/model selection despite known valid values. Error handling descriptions absent, tools do not guide recovery. Composition is reasonable (single responsibility per tool), but chaining IDs are not guaranteed in responses.
Create a new agent with specified configuration
Delete an agent by ID
Disable an agent
Enable an agent
Get detailed information about an agent
List all agents
Retrieve logs for an agent
List available AI models for agent configuration
Output schemas undocumented. LLMs cannot predict response structure for any of 39 tools. Responses must include chaining IDs (e.g., agent_id after agent_add) but this is not guaranteed or documented.
Parameter descriptions lack format constraints. 'expires_at' accepts string but does not specify RFC3339 format. 'path' parameters lack validation rules (absolute vs relative, existence checks). LLMs cannot infer valid input formats.
No enum constraints on platform/model selection. agent_add accepts 'platform' as free-form string (e.g., 'anthropic', 'openai') but valid values are not declared. agent_models returns available models but agent_add does not enforce them.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | <=2025-11-25 | v2 |
Report progress or status for a scheduled task
Manually trigger an agent execution
Update an existing agent configuration
Get daemon health and status information
Get available tools filtered by scope
List stored intelligence entries (facts and patterns)
Store or update a fact, pattern, or session summary
Create a new loop (workflow automation)
Archive a completed loop
Resume a paused loop
Delete a loop
Get detailed information about a loop
List all loops
Pause a running loop
Reset a loop to initial state
Manually trigger a loop execution
Update loop configuration
Ingest documents into the RAG system
Query the RAG system for relevant documents
Create a new standalone spec
Delete a spec
Get detailed information about a spec
List all specs
Update a spec
Broadcast a message to all agents in the session
Declare intent to work on a file or task
Get workspace context and detect conflicts
Report final status and close session
Create a file watcher
Delete a file watcher
List all file watchers
Destructive tools (agent_delete, loop_delete, spec_delete, watch_delete) lack confirmation or dry-run support. No error recovery guidance. LLMs cannot know whether to retry or escalate on failure.
List tools (agent_list, loop_list, spec_list, watch_list) lack pagination parameters (limit, offset, cursor). No documented result count or truncation behavior. Large result sets will blow context windows.