Loading ToolBench
AI agent orchestration in s-expressions. A DSL and runtime for composing LLM-powered workflows with support for sequential pipelines, parallel fan-out, conditionals, and integration with external services via plugins.
AgentScript exposes 32 tools via a custom STDIO-only DSL, not as MCP-compliant tool definitions. Tool names are present but lack formal JSON Schema registration. Descriptions are brief (10-50 chars) and generic. No visible input parameter schemas with type definitions, descriptions, or constraints. Output schemas are undocumented. Error handling is absent from tool definitions. Security patterns (secret injection, permission gates, audit trails) are not evident. Tools like 'exec' (DESTRUCTIVE) and 'agent'/'mcpagent' (WRITE) lack confirmation or dry-run patterns. The codebase is a DSL interpreter, not an MCP server, tool definitions are inferred from plugin file names and brief help text, not explicit schema registration.
Run natural language instructions as a DSL pipeline via Claude
Cache results of expensive operations
Deploy and schedule DSL scripts as Google Cloud Run Jobs
Get cryptocurrency price and market data
Render data as self-contained interactive HTML tables
Send email messages
Execute shell commands
Deploy HTML to GitHub Pages
No JSON Schema input definitions visible for any tool. All 32 tools lack formal parameter type definitions, constraints, and descriptions required by MCP spec.
Tool descriptions are generic and under 50 characters. LLMs cannot distinguish between similar tools (e.g., 'agent' vs 'mcpagent' vs 'plugagent') or understand when to invoke them.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 30 | 2026-07-28+ | v2 |
Run inference on HuggingFace models
Search for job listings
Knowledge graph operations using Apache AGE and pgvector
Call tools from connected MCP servers
AI-driven tool selection over connected MCP servers using Claude or Gemini
Search the official MCP registry for available servers
Network diagnostics and connectivity checks
Fetch news articles on a topic
Send notifications via Slack, email, or other channels
Run local LLM inference using Ollama
Fill PDF forms using AI-powered field detection
Search and answer questions using Perplexity AI
Generate new plugins from English descriptions using Claude
Vector search and retrieval-augmented generation using PostgreSQL and pgvector
Fetch posts and comments from Reddit
Code review forum with Claude, Gemini, and GPT-4 debate
Fetch and parse RSS feeds
Save output to a file
Search the web using SerpAPI or similar search service
Get stock price and market data for a ticker symbol
Summarize text or web content using an LLM
Fetch tweets and search Twitter
Get current weather information for a location
Send WhatsApp messages via Twilio
Destructive tools (exec, delete operations) and write tools (email, whatsapp, notify, cloudrun) lack confirmation, dry-run, or permission-gate patterns. No audit trail or error recovery guidance.
No output schemas documented. Agents cannot plan downstream tool calls or extract required fields (e.g., IDs for chaining). Response structure is opaque.
No error handling patterns visible. Tools do not guide LLMs on recovery (retryable vs fatal), actionable error messages, or alternative paths when resources are not found.